Patent Application
20080244239
1. Field of the Invention
The present invention relates generally to an improved data processing system. In particular, the present invention provides a method and apparatus for obtaining performance data in a data processing system. Still more particularly, the present invention provides a method and apparatus for hardware assistance to software tools in obtaining performance data in a data processing system.
2. Description of Related Art
In analyzing and enhancing performance of a data processing system and the applications executing within the data processing system, it is helpful to know which software modules within a data processing system are using system resources. Effective management and enhancement of data processing systems requires knowing how and when various system resources are being used. Performance tools are used to monitor and examine a data processing system to determine resource consumption as various software applications are executing within the data processing system. For example, a performance tool may identify the most frequently executed modules and instructions in a data processing system, or may identify those modules which allocate the largest amount of memory or perform the most I/O requests. Hardware performance tools may be built into the system or added at a later point in time.
One known software performance tool is a trace tool. A trace tool may use more than one technique to provide trace information that indicates execution flows for an executing program. One technique keeps track of particular sequences of instructions by logging certain events as they occur, a so-called event-based profiling technique. For example, a trace tool may log every entry into, and every exit from, a module, subroutine, method, function, or system component. Alternately, a trace tool may log the requester and the amounts of memory allocated for each memory allocation request. Typically, a time-stamped record is produced for each such event. Corresponding pairs of records, similar to entry-exit records, also are used to trace execution of arbitrary code segments, starting and completing I/O or data transmission, and for many other events of interest.
In order to improve performance of code generated by various families of computers, it is often necessary to determine where time is being spent by the processor in executing code, such efforts being commonly known in the computer processing arts as locating “hot spots”. Ideally, one would like to isolate such hot spots at the instruction and/or source line of code level in order to focus attention on areas which might benefit most from improvements to the code.
Another trace technique involves periodically sampling a program's execution flows to identify certain locations in the program in which the program appears to spend large amounts of time. This technique is based on the idea of periodically interrupting the application or data processing system execution at regular intervals, so-called sample-based profiling. At each interruption, information is recorded for a predetermined length of time or for a predetermined number of events of interest. For example, the program counter of the currently executing thread, which is an executable portion of the larger program being profiled, may be recorded during the intervals. These values may be resolved against a load map and symbol table information for the data processing system at post-processing time, and a profile of where the time is being spent may be obtained from this analysis.
However, creating tools such as these to find answers related to specific situations or problems can take much effort and can be very difficult to calibrate as the software tools themselves affect the system under test. Moreover, the application is modified by a software developer during a post-processing phase in accordance with the insights that have been gained through a post-processing analysis of the information that was gather during the execution of an application.
Therefore, it would be advantageous to have an improved method, apparatus, and computer program product for providing hardware assistance for gathering performance information that significantly reduces the overhead in gathering such information and that significantly reduces the impact on the application that is being profiled. Furthermore, it would be advantageous to allow a software application during a runtime phase to use the information that is dynamically gathered from the assisting hardware in order to autonomically affect the behavior of the software application.
A method, an apparatus, and a computer program product in a data processing system are presented for using hardware assistance for gathering performance information that significantly reduces the overhead in gathering such information. Performance indicators are associated with instructions or memory locations, and processing of the performance indicators enables counting of events associated with execution of those instructions or events associated with accesses to those memory locations. The performance information that has been dynamically gathered from the assisting hardware is available to the software application during runtime in order to autonomically affect the behavior of the software application, particularly to enhance its performance. For example, the counted events may be used to autonomically collecting statistical information about the ability of a software application to successfully acquire a semaphore.
The novel features believed characteristic of the invention are set forth in the appended claims. The invention itself, further objectives, and advantages thereof, will be best understood by reference to the following detailed description when read in conjunction with the accompanying drawings, wherein:
With reference now to
An operating system runs on processor 102 and is used to coordinate and provide control of various components within data processing system 100 in
Those of ordinary skill in the art will appreciate that the hardware in
For example, client 100, if optionally configured as a network computer, may not include SCSI host bus adapter 112, hard disk drive 126, tape drive 128, and CD-ROM 130. In that case, the computer, to be properly called a client computer, includes some type of network communication interface, such as LAN adapter 110, modem 122, or the like. As another example, client 100 may be a stand-alone system configured to be bootable without relying on some type of network communication interface, whether or not client 100 comprises some type of network communication interface. As a further example, client 100 may be a personal digital assistant (PDA), which is configured with ROM and/or flash ROM to provide non-volatile memory for storing operating system files and/or user-generated data. The depicted example in
The processes of the present invention are performed by processor 102 using computer implemented instructions, which may be located in a memory such as, for example, main memory 104, memory 124, or in one or more peripheral devices 126-130.
Turning next to
In a preferred embodiment, processor 210 is a single integrated circuit superscalar microprocessor. Accordingly, as discussed further herein below, processor 210 includes various units, registers, buffers, memories, and other sections, all of which are formed by integrated circuitry. Also, in the preferred embodiment, processor 210 operates according to reduced instruction set computer (“RISC”) techniques. As shown in
BIU 212 is connected to an instruction cache 214 and to data cache 216 of processor 210. Instruction cache 214 outputs instructions to sequencer unit 218. In response to such instructions from instruction cache 214, sequencer unit 218 selectively outputs instructions to other execution circuitry of processor 210.
In addition to sequencer unit 218, in the preferred embodiment, the execution circuitry of processor 210 includes multiple execution units, namely a branch unit 220, a fixed-point unit A (“FXUA”) 222, a fixed-point unit B (“FXUB”) 224, a complex fixed-point unit (“CFXU”) 226, a load/store unit (“LSU”) 228, and a floating-point unit (“FPU”) 230. FXUA 222, FXUB 224, CFXU 226, and LSU 228 input their source operand information from general-purpose architectural registers (“GPRs”) 232 and fixed-point rename buffers 234. Moreover, FXUA 222 and FXUB 224 input a “carry bit” from a carry bit (“CA”) register 239. FXUA 222, FXUB 224, CFXU 226, and LSU 228 output results (destination operand information) of their operations for storage at selected entries in fixed-point rename buffers 234. Also, CFXU 226 inputs and outputs source operand information and destination operand information to and from special-purpose register processing unit (“SPR unit”) 237.
FPU 230 inputs its source operand information from floating-point architectural registers (“FPRs”) 236 and floating-point rename buffers 238. FPU 230 outputs results (destination operand information) of its operation for storage at selected entries in floating-point rename buffers 238.
In response to a Load instruction, LSU 228 inputs information from data cache 216 and copies such information to selected ones of rename buffers 234 and 238. If such information is not stored in data cache 216, then data cache 216 inputs (through BIU 212 and system bus 211) such information from a system memory 239 connected to system bus 211. Moreover, data cache 216 is able to output (through BIU 212 and system bus 211) information from data cache 216 to system memory 239 connected to system bus 211. In response to a Store instruction, LSU 228 inputs information from a selected one of GPRs 232 and FPRs 236 and copies such information to data cache 216.
Sequencer unit 218 inputs and outputs information to and from GPRs 232 and FPRs 236. From sequencer unit 218, branch unit 220 inputs instructions and signals indicating a present state of processor 210. In response to such instructions and signals, branch unit 220 outputs (to sequencer unit 218) signals indicating suitable memory addresses storing a sequence of instructions for execution by processor 210. In response to such signals from branch unit 220, sequencer unit 218 inputs the indicated sequence of instructions from instruction cache 214. If one or more of the sequence of instructions is not stored in instruction cache 214, then instruction cache 214 inputs (through BIU 212 and system bus 211) such instructions from system memory 239 connected to system bus 211.
In response to the instructions input from instruction cache 214, sequencer unit 218 selectively dispatches the instructions to selected ones of execution units 220, 222, 224, 226, 228, and 230. Each execution unit executes one or more instructions of a particular class of instructions. For example, FXUA 222 and FXUB 224 execute a first class of fixed-point mathematical operations on source operands, such as addition, subtraction, ANDing, ORing and XORing. CFXU 226 executes a second class of fixed-point operations on source operands, such as fixed-point multiplication and division. FPU 230 executes floating-point operations on source operands, such as floating-point multiplication and division.
As information is stored at a selected one of rename buffers 234, such information is associated with a storage location (e.g. one of GPRs 232 or carry bit (CA) register 242) as specified by the instruction for which the selected rename buffer is allocated. Information stored at a selected one of rename buffers 234 is copied to its associated one of GPRs 232 (or CA register 242) in response to signals from sequencer unit 218. Sequencer unit 218 directs such copying of information stored at a selected one of rename buffers 234 in response to “completing” the instruction that generated the information. Such copying is called “writeback.”
As information is stored at a selected one of rename buffers 238, such information is associated with one of FPRs 236. Information stored at a selected one of rename buffers 238 is copied to its associated one of FPRs 236 in response to signals from sequencer unit 218. Sequencer unit 218 directs such copying of information stored at a selected one of rename buffers 238 in response to “completing” the instruction that generated the information.
Processor 210 achieves high performance by processing multiple instructions simultaneously at various ones of execution units 220, 222, 224, 226, 228, and 230. Accordingly, each instruction is processed as a sequence of stages, each being executable in parallel with stages of other instructions. Such a technique is called “pipelining.” In a significant aspect of the illustrative embodiment, an instruction is normally processed as six stages, namely fetch, decode, dispatch, execute, completion, and writeback.
In the fetch stage, sequencer unit 218 selectively inputs (from instruction cache 214) one or more instructions from one or more memory addresses storing the sequence of instructions discussed further hereinabove in connection with branch unit 220, and sequencer unit 218.
In the decode stage, sequencer unit 218 decodes up to four fetched instructions.
In the dispatch stage, sequencer unit 218 selectively dispatches up to four decoded instructions to selected (in response to the decoding in the decode stage) ones of execution units 220, 222, 224, 226, 228, and 230 after reserving rename buffer entries for the dispatched instructions' results (destination operand information). In the dispatch stage, operand information is supplied to the selected execution units for dispatched instructions. Processor 210 dispatches instructions in order of their programmed sequence.
In the execute stage, execution units execute their dispatched instructions and output results (destination operand information) of their operations for storage at selected entries in rename buffers 234 and rename buffers 238 as discussed further hereinabove. In this manner, processor 210 is able to execute instructions out-of-order relative to their programmed sequence.
In the completion stage, sequencer unit 218 indicates an instruction is “complete.” Processor 210 “completes” instructions in order of their programmed sequence.
In the writeback stage, sequencer 218 directs the copying of information from rename buffers 234 and 238 to GPRs 232 and FPRs 236, respectively. Sequencer unit 218 directs such copying of information stored at a selected rename buffer. Likewise, in the writeback stage of a particular instruction, processor 210 updates its architectural states in response to the particular instruction. Processor 210 processes the respective “writeback” stages of instructions in order of their programmed sequence. Processor 210 advantageously merges an instruction's completion stage and writeback stage in specified situations.
In the illustrative embodiment, each instruction requires one machine cycle to complete each of the stages of instruction processing. Nevertheless, some instructions (e.g., complex fixed-point instructions executed by CFXU 226) may require more than one cycle. Accordingly, a variable delay may occur between a particular instruction's execution and completion stages in response to the variation in time required for completion of preceding instructions.
Completion buffer 248 is provided within sequencer 218 to track the completion of the multiple instructions which are being executed within the execution units. Upon an indication that an instruction or a group of instructions have been completed successfully, in an application specified sequential order, completion buffer 248 may be utilized to initiate the transfer of the results of those completed instructions to the associated general-purpose registers.
In addition, processor 210 also includes performance monitor unit 240, which is connected to instruction cache 214 as well as other units in processor 210. Operation of processor 210 can be monitored utilizing performance monitor unit 240, which in this illustrative embodiment is a software-accessible mechanism capable of providing detailed information descriptive of the utilization of instruction execution resources and storage control. Although not illustrated in
Performance monitor unit 240 includes an implementation-dependent number (e.g., 2-8) of counters 241-242, labeled PMC1 and PMC2, which are utilized to count occurrences of selected events. Performance monitor unit 240 further includes at least one monitor mode control register (MMCR). In this example, two control registers, MMCRs 243 and 244 are present that specify the function of counters 241-242. Counters 241-242 and MMCRs 243-244 are preferably implemented as SPRs that are accessible for read or write via MFSPR (move from SPR) and MTSPR (move to SPR) instructions executable by CFXU 226. However, in one alternative embodiment, counters 241-242 and MMCRs 243-244 may be implemented simply as addresses in I/O space. In another alternative embodiment, the control registers and counters may be accessed indirectly via an index register. This embodiment is implemented in the IA-64 architecture in processors from Intel Corporation.
Additionally, processor 210 also includes interrupt unit 250, which is connected to instruction cache 214. Additionally, although not shown in
The present invention provides an ability to monitor the execution of specific instructions as well as the access of specific memory locations during the execution of a program. Specifically, a spare field may be used to hold an indicator that identifies the instruction or memory location as one that is to be monitored by a performance monitor unit or by some other unit in a processor. Alternatively, the indicator may be stored in another location in association with the instruction or memory location. In the case in which the indicator is placed in the instruction, a spare field is typically used, but in some cases the instruction may be extended to include the space needed for the indicator. With this case, the architecture of the processor may require changes. For example, a 64 bit architecture may be changed to a 65 bit architecture to accommodate the indicator. With respect to accesses of data, an indicator may be associated with the data or memory locations in which the data is located.
Turning now to
As part of this processing of instructions, instruction cache 300 determines which instructions are associated with indicators. These indicators also are referred to as “performance indicators” in these examples. Signals 304 have been associated with performance indicators. As a result, signals 304 for the instructions are sent to performance monitor unit 306. Performance monitor unit 306 is an example of performance monitor unit 240 in
When instruction cache 300 determines that an instruction associated with an indicator is present, a signal is sent to indicate that a marked instruction is being executed. In these examples, a marked instruction is an instruction associated with a performance indicator. Alternatively, a performance indicator may indicate that all items or instructions in a bundle are marked to be counted. Additionally, signals for these instructions are sent by instruction cache 300 to the appropriate functional unit. Depending on the particular implementation, a functional unit other than performance monitor unit 306 may count execution of instructions. In the case that the performance indicators are in the instructions, or in the bundles, the cache unit, instruction cache 300, detects the indicators and sends signals to performance monitor unit 306.
When signals for these instructions are received by performance monitor unit 306, performance monitor unit 306 counts events associated with execution of instructions 304. As illustrated, performance monitor unit 306 is programmed only to count events for instructions associated with performance indicators. In other words, an indicator associated with a instruction or memory location is used to enable counting of events associated with the instruction or memory location by performance monitor unit 306. If an instruction is received by instruction cache 300 without a performance indicator, then events associated with that instruction are not counted. In summary, the performance indicators enable the counting on a per instruction or per memory location basis in a processor.
Performance monitor unit 306 counts events for instructions associated with performance indicators, if performance monitor unit 306 is set in a mode to count metrics enabled for these types of marked instructions. In some cases, performance monitor unit 306 may be set to perform some other type of counting, such as counting execution of all instructions, which is a currently available function.
With respect to the accessing of data in memory locations, the data and indicators are processed by a data cache, such as data cache 216 in
Turning next to
When processor 400 receives an instruction from cache 402, processor 400 checks performance instrumentation shadow cache 404 to see whether a performance indicator is associated with the instruction. A similar check is made with respect to accesses of memory locations containing data. In one embodiment, a full shadow word is provided for each corresponding word that does not affect the actual data segments. In other words, processor 400 allows for the architecture or configuration of cache 402 to remain unchanged. In these examples, the mapping described is word for word. However, some other type of mapping may be used, such as a shadow bit per data word in which a bit in performance instrumentation shadow cache 404 corresponds to one word of data.
With respect to this type of architecture, the compilers, using this feature, create the debug information in a separate work area from the data area themselves in a manner similar to debug symbols. When a module is loaded, the extra information, performance indicators, is prepared by the loader so that it will be available to incorporate into performance instrumentation shadow cache 404 when instructions are loaded into cache 402. These cache areas may be intermingled and either marked as such or understood by the mode of operation. Processor 400 uses the performance indicators to determine how the related data accesses and instruction executions are to be counted or made to take exceptions. In these examples, the process is programmed by a debugger or a performance analysis program to know whether to use the shadow information while it is executing instructions.
Turning next to
Spare bits within bundle 500 are used to hold indicators of the present invention. For example, indicators 510, 512, and 514 are located within instruction slots 502, 504, and 506, respectively. These indicators may take various forms and may take various sizes depending on the particular implementation. Indicators may use a single bit or may use multiple bits. A single bit may be used to indicate that events are to be counted in response to execution of that instruction. Multiple bits may be used to identify a threshold, such as a number of processor or clock cycles for instruction execution that may pass before events should be counted. Further, these bits may even be used as a counter for a particular instruction. A similar use of fields may be used for indicators that mark data or memory locations.
Alternatively, template 508 may be used to contain a bundle of related indicators, so that one bit is used to identify all of the instructions in a bundle. Also, the bundle itself could be extended to be 256 bits or some other number of bits to contain the extra information for the performance indicators.
Turning next to
Data or memory locations containing data may be marked with indicators in a similar manner. These indicators are used in counting accesses to the data or memory locations in these examples. In
Turning now to
The process begins by receiving a bundle (step 700). In these examples, each bundle has a format similar to bundle 500 in
If a performance indicator is present, a signal is sent to a performance monitor unit (step 706). Upon receiving this signal, the performance monitor unit will count events associated with the execution of the instruction. Additionally, the instruction is processed (step 708). Processing of the instruction includes, for example, sending the instruction to the appropriate functional unit for execution.
Thereafter, a determination is made as to whether additional unprocessed instructions are present in the bundle (step 710). If additional unprocessed instructions are present in the bundle, the process returns to step 702 as described above. Otherwise, the process terminates. Turning back to step 704, if the performance indicator is not present, the process proceeds directly to step 708.
Turning now to
The process begins by receiving a bundle (step 800). An instruction in the bundle is identified (step 802). A determination is made as to whether a performance indicator associated with the instruction is present (step 804). The signal sent to the interrupt unit to indicate an exception is to be raised is different from the signal sent to the performance monitor unit. For example, an instruction may be associated with a specific performance indicator having a first value that causes a signal to be sent to the interrupt unit. A second value for a performance indicator may be used to send a different signal to the performance monitor unit. If a performance indicator having the first value is present, the signal is sent to an interrupt unit (step 806). Upon receiving this signal, the interrupt unit initiates appropriate call flow support to process this interrupt. The call flow support may, for example, record cache misses that may be missed by a functional unit trying to access instructions or data in a cache.
Additionally, the instruction is processed (step 808). Processing of the instruction includes, for example, sending the instruction to the appropriate functional unit for execution.
Thereafter, a determination is made as to whether additional unprocessed instructions are present in the bundle (step 810). If additional unprocessed instructions are present in the bundle, the process returns to step 802 as described above. Otherwise, the process terminates. Turning back to step 804, if the performance indicator is not present, the process proceeds directly to step 808.
With reference now to
The process begins by identifying a request to access a memory location (step 900). In response to identifying this request, a determination is made as to whether a performance indicator is associated with the memory location (step 902). If a performance indicator is associated with the memory location, an interrupt is generated by sending a signal to the interrupt unit (step 904). Thereafter, the access to the memory location is processed (step 906) with the process terminating thereafter.
In
The process begins by receiving a signal from an instruction cache indicating that an instruction with a performance indicator is being processed (step 1000). Next, events associated with the instruction being processed are counted (step 1002) with the process terminating thereafter. The counting of events may be stored in a counter, such as counter 241 in
With reference next to
The process begins by determining whether an instruction associated with a performance indicator has been received (step 1100). In this example, the indicator causes counting of events for this instruction and all subsequent instructions executed by the processor. Alternatively, the indicator could be an instruction itself which indicates the new mode of counting is to be started. If an instruction with an indicator has been received, a flag is set to start counting events for instructions (step 1102). This flag indicates that counting events for instructions should start.
Next, a determination is made as to whether an instruction with an indicator has been received (step 1104). Alternatively, the indicator could be an instruction itself which indicates the new mode of counting is to be stopped. If an instruction with an indicator is received, the flag is unset to stop counting the events (step 1106) with the process terminating thereafter.
The indicator in step 1100 and step 1104 may be the same indicator in which the indicator toggles the setting and unsetting of the flag. In another implementation, two different indicators may be used in which a first indicator only sets the flag. A second indicator is used to unset the flag. Communication between a cache unit, such as an instruction cache or a data cache, and the performance monitor unit to indicate a mode of counting may be implemented simply with a high signal when counting is to occur and a low signal when counting is no longer enabled.
With reference next to
The process begins by checking a flag (step 1200). A determination is made as to whether the flag is set (step 1202). If the flag is set, a signal is sent to the performance monitor unit to enable this unit to count events (step 1204) with the process terminating thereafter. Otherwise, a signal is sent to the performance monitor unit to disable the counting of events (step 1206) with the process terminating thereafter.
The processes illustrated in
Turning now to
The process begins by receiving an instruction associated with a performance indicator (step 1300). A threshold is identified for the instruction (step 1302). In these examples, the threshold relates to a number of processor or clock cycles needed to complete an instruction. If the cache latency or amount of time needed to access the cache exceeds the threshold value, that event is counted. The threshold value is set within the indicator in these examples.
For example, three bits may be used to set eight different values for the threshold. For example, “xx1”=10 cycles, “x1x”=50 cycles, and “1xx”=100 cycles. Some combination of these three bits may be used to set values for the threshold. More or fewer bits may be used and different values may be assigned to the bits depending on the specific implementation. The meaning of the bits may also be controlled through an interface, such as a set of registers that may be used to set the meaning of each of the bits. These registers are ones that are added to the processor architecture for this specific purpose.
Cycles for executing the instruction are monitored (step 1304). A determination is made as to whether the threshold has been exceeded for this instruction (step 1306). If the threshold has been exceeded, then a selected action is performed (step 1308). This selected action may take different forms depending on the particular implementation. For example, a counter may be incremented each time the threshold is exceeded. Alternatively, an interrupt may be generated. The interrupt may pass control to another process to gather data. For example, this data may include a call stack and information about the call stack. A stack is a region of reserved memory in which a program or programs store status data, such as procedure and function call addresses, passed parameters, performance monitor counter values, and sometimes local variables.
A determination is made as to whether monitoring is to end (step 1310). Step 1310 may be implemented one instruction at a time. When an instruction is executed or the threshold is exceeded, a signal is sent. In this example, execution of a single instruction results in one signal being sent. In the case in which multiple instructions may be executed at the same time, multiple signals may be needed to indicate the execution of each instruction. In some embodiments, a sampling approach may be supported, where the threshold is only supported for one instruction at a time. This may be done by only supporting thresholds for those instructions that are in a particular position in the processor's instruction queue. In other embodiments, one signal may be sent if at least one of the marked instructions exceeds the threshold. For each instruction in which a threshold is exceeded, a separate signal is raised or generated for that instruction.
If the monitoring is to end, the collected information is sent to a monitoring program (step 1312), with the process terminating thereafter. Otherwise, the process returns to step 1304 as described above. In step 1306, if the threshold is not exceeded for the instruction, the process proceeds directly to step 1310.
A similar process may be implemented in a data cache, such as data cache 216 in
As with the other examples, these indicators may be included as part of the instruction or with the data in a memory location. Alternatively, these indicators may be found in a performance instrumentation shadow cache or memory in association with the instruction or data.
With reference to
The process begins by receiving data associated with a performance indicator (step 1400). A determination is made as to whether a memory location for the data has been accessed (step 1402). If the memory location has been accessed, then a counter is incremented (step 1404). A determination is made as to whether monitoring is to end (step 1406). If monitoring of the memory location is to end, the process terminates. Otherwise, the process returns to step 1402. In step 1402, if the memory location is not accessed, then the process proceeds to step 1406.
Turning to
In one embodiment the format simply has a performance instrumentation shadow cache entry for each of its block or sector references and moves meta data 1504 to its corresponding shadow entry or entries. Instead of having a performance instrumentation shadow cache, the internal format of the cache itself may be modified to contain meta data 1504. In embodiments where the instruction stream itself is modified to contain the meta data, then either the loader updates the instruction stream to contain the appropriate indicators and work areas or compiler 1500 has generated the code to contain meta data 1504. In either case, after the code is loaded, the processor receives the meta data 1504.
In addition, meta data 1504 may be placed into performance instrumentation shadow memory 1505 in association with instructions 1502. Compiler 1500 produces information in a table or debug data section. The performance monitoring program loads this information into shadow data areas in performance instrumentation shadow memory 1505. Alternatively, the debug areas may be automatically populated by the operating system and the processor working together.
Instructions 1502 may then be executed by processor 1508. Compiler 1500 may set a register such as mode register 1510 in processor 1508. When this register is set, processor 1508 looks at meta data 1504 in performance instrumentation shadow memory 1505 when executing instructions 1502 to determine whether performance indicators in meta data 1504 are associated with instructions that are being executed in instructions 1502. These performance indicators are handled using processes, such as those described above with reference to
A similar process may be performed with respect to data in memory location 1512. Depending on the particular implementation, meta data 1504 may be placed within the instruction or within the data, rather than in performance instrumentation shadow memory 1505. However, by placing meta data 1504 in performance instrumentation shadow memory 1505, the generation of meta data 1504 may be performed dynamically when meta data 1504 is placed in performance instrumentation shadow memory 1505.
This feature allows for selection and monitoring of instructions to occur without having to modify the program. In other words, compiler 1500 may generate meta data 1504 after instructions 1502 have been compiled for execution by processor 1508. Setting mode register 1510 causes processor 1508 to look for meta data 1504 in performance instrumentation shadow memory 1505 without having to modify instructions 1502. In these examples, meta data 1504 take the form of performance indicators that tell processor 1508 how to handle the execution of instructions 1502 and/or data accesses to memory location 1512.
Turning next to
In this example, meta data 1600 includes 5 entries, entry 1602, 1604, 1606, 1608, and 1610 as indicated by line 1612 in meta data 1600. Each of these entries includes an offset, a length, and a flag for describing the instrumentation of code in this example.
Entry 1602 has an offset of 0 with an entry length of 120 bytes. Flag 1614 indicates that all instructions within the range indicated by entry length 1616 need to be counted. In these examples, each instruction has a length of 4 bytes. Entry 1604 has an entry length of 4 bytes, which corresponds to an instruction. Flag 1618 indicates that an exception should be generated upon execution of this instruction.
In entry 1606, an instruction beginning at an offset of 160 bytes is associated with flag 1620. This flag indicates that the instruction should be counted if the threshold, 100 cycles, is exceeded.
Flag 1622 in entry 1608 indicates that tracing should start at the instruction having an offset of 256 bytes. Tracing stops as indicated by flag 1624 in entry 1610, which has a flag for the instruction at an offset of 512 bytes.
These flags are used to generate the performance indicators that are associated with the instructions. The operating system moves this meta data generated by the compiler and processes the meta data into a performance instrumentation shadow memory, such as performance instrumentation shadow memory 1506 in FIG. 15. Alternatively, this meta data may be placed into fields within the instructions depending on the particular implementation.
With reference now to
At program compile time, the compiler generates a new performance instrumentation data section as previously described. At program load time, the loader queries the processor to determine cache line size. The loader parses perfinst segment 1728 and constructs a shadow segment, in the format required by the processor, for any text or data segment that the loader loads. This shadow segment is placed into new performance instrumentation shadow cache 1730.
Each block in the shadow segment contains meta data for instructions or data in the corresponding primary cache block. This meta data includes, for example, flags, tag fields, threshold, and count fields for each tagged item in a block in primary segment 1702. This meta data also may include a flag that represents all the instructions or data in the block.
The loader constructs a table mapping, translation table 1726, for each block in primary segment 1702 to a corresponding perfinst block, such as block 1732, 1734, 1736, 1738, 1740, 1742, 1744, 1746, 1748, 1750, and 1752 in perfinst segment 1728. Further, the loader registers the head of this table, translation table 1726, and the location and size of primary segment 1702 with the processor.
At page replacement time, paging software provides a new interface to associate perfinst segment 1728 with the corresponding primary segment, primary segment 1702. When primary segment 1702 pages in or out, perfinst segment 1728 pages in or out as well.
At cache line replacement time, the processor contains new performance instrumentation shadow cache 1730 with cache frames directly associated with the frames in the existing data and instruction caches, such as existing cache 1700. When the processor's instruction or data cache loads a new line, the cache also must load the corresponding perfinst block into the performance instrumentation shadow cache, new performance instrumentation shadow cache 1730. The processor sees (from the registration data given by the loader at program load time) that the processor is bringing a block into its cache that has an associated perfinst segment, perfinst segment 1728. The processor looks in translation table 1726 associated with this segment, finds a reference to the perfinst block corresponding to the block it is about to load and loads the perfinst block into new performance instrumentation shadow cache 1730. In these examples, cache misses associated with meta data are not signaled or are treated differently from cache misses associated data in a primary cache block, such as in primary segment 1702.
With reference now to
The process begins by identifying an instruction for profiling (step 1800). This instruction may be, for example, one that has been executed more than a selected number of times. Meta data is generated for the identified instruction (step 1802). This meta data takes the form of a performance indicator. The performance indicator may, for example, increment a counter each time the instruction is executed, increment a counter if the number of cycles needed to execute the instruction exceeds a threshold value, toggle counting of events for all instructions for all events after this instruction, or count events occurring in response to executing the instruction. In a preferred embodiment, the counters are in the associated performance instrumentation shadow cache and take some number of bits to allow for a one to one correspondence between the data or instructions in the cache and the bits reserved for counting.
The meta data is then associated with the instruction (step 1804). Next, a determination is made as to whether more instructions are present for processing (step 1806). If additional instructions are present, the process returns to step 1800. Otherwise, the process terminates. A similar process may be used to dynamically generate meta data for data in memory locations.
With reference now to
The process begins by identifying a memory location for profiling (step 1900). Step 1900 occurs by detecting access to a marked location. Meta data is generated for the identified memory location (step 1902). This meta data takes the form of a performance indicator. The performance indicator may, for example, increment a counter each time the memory location is accessed, increment a counter if the number of cycles needed to access the memory location exceeds a threshold value, or toggle counting of all accesses to memory locations. The meta data is then associated with the memory location (step 1904). Next, a determination is made as to whether more memory locations are present for processing (step 1906). If additional memory locations are present, the process returns to step 1900. Otherwise, the process terminates.
Turning now to
The process begins by executing an instruction (step 2000). A determination is made as to whether a counter is associated with the instruction (step 2002). The counter may be included in a field within the instruction or may be in a performance instrumentation shadow memory. If a counter is associated with the instruction, the counter is incremented (step 2004) with the process terminating thereafter. Otherwise, the process terminates without incrementing the counter. The counter may be reset if the counter exceeds a threshold value.
When the counter is implemented as part of the instructions, the counter may be of limited size. In this case, a threshold value for the counter may be set to indicate when the counter is in danger of overflowing. The counter may then be reset after the value has been read. This value may be read by a performance monitor unit or by a program used to analyze data. APIs may be implemented to access this data.
Turning now to
The process begins by detecting access to a memory location (step 2100). A determination is made as to whether a counter is associated with the memory location (step 2102). The counter may be included within the memory location or may be in a performance instrumentation shadow memory. If a counter is associated with the memory location, the counter is incremented (step 2104) with the process terminating thereafter. Otherwise, the process terminates without incrementing the counter.
With reference next to
When the instruction or data cache pages are loaded into memory, the operating system program loader/linker and/or the performance monitoring program, reads the meta data generated by the compiler and determines that counting is associated with instruction or data access, then the loading process allocates data areas to maintain the counters as part of its perfinst segment. The size of the counters and the granularity of the data access determine the amount of work area to be allocated.
In a simple case, the granularity of the data or instruction access could be word size (so that an access to any byte in the word is considered an access) and the counts could also be a word size. In this case, one to many mapping is present between the primary segment and the perfinst segment (a full word to contain the counts or threshold is not required). The loading process allocates a shadow page or pages and tells the processor to use the shadow page(s) to contain the counts. Details of this mapping are described above with reference to
In an alternative embodiment, the compiler allocates the work areas to maintain the counts and indicates the placement of these work areas in its generated data areas. An entry in the meta data could indicate the start of the data, the number of bytes of data, granularity of the data, the start of the count area, and the granularity of each counting unit. In either case, the meta data is loaded into the processor and the processor populates its internal (shadow) cache with the meta data. In illustrative embodiments in which the instruction stream itself is modified to contain the meta data, then either the loader updates the instruction stream to contain the appropriate indicators and work areas or the compiler has generated the code to contain the meta data. In either case, after the code is loaded, the processor receives the meta data.
Data unit 2206 may be implemented as data cache 206 in
Each of these events, instruction execution and data access, results in incrementing of a counter. The mechanism of the present invention provides an interface, hardware interface 2212, to access this collected data. In these examples, hardware interface 2212 takes the form of an application programming interface (API) for operating system 2214. In this way, analysis tool 2216 may obtain data from counter 2204 and counter 2210. Analysis tool 2216 may take many forms, such as for example, Oprofile, which is a known system wide profiler for Linux systems. Although the examples in
In
Alternatively, a fixed number of counts of a performance monitor counter may be used instead of a timer. This program profiles subroutines that are used to indicate where time is spent within a program. A program having usage over a certain threshold also is referred to as being “hot”. By using information from profiler 2300, routines of interest, such as subroutine 2304 in program 2302 may be identified.
With this information, the instructions in subroutine 2304 may be autonomically modified by analysis tool 2306 to allow counting of the execution of subroutine 2304. Additional routines may be identified for modification by analysis tool 2306. For example, subroutine 2304 also may be identified as a routine of interest with the instructions of this routine being modified to allow counting of the execution of subroutine 2304. The modification of the code in these routines includes associating performance indicators with one or more instructions within each of these subroutines.
After the instructions in these routines have been modified by analysis tool 2306, program 2302 is then executed by processor 2308. Processor 2308 executes program 2302 and provides counts for these routines. For example, the counting of instructions executed and the number of cycles used in executing a routine may be performed by processor 2308 using the mechanisms described above.
With reference to
The process begins by identifying instructions of interest using data from a profiler (step 2400). This profiler may be, for example, a timer profiler found in AIX. An instruction from the identified instructions is selected for modification (step 2402). Thereafter, a performance indicator is dynamically added to the selected instruction (step 2404).
In step 2404, the instruction may be added in a manner such that the instructions do not need to be modified for execution. A performance instrumentation shadow memory, such as performance instrumentation shadow memory 1506 in
A determination is then made as to whether additional identified instructions are present for modification (step 2406). If additional instructions are present for modification, the process returns to step 2402. Otherwise, the process terminates.
Turning next to
In this example, program 2500 contains three pages, page 2502, page 2504, and page 2506. Scanning daemon 2508 associates performance indicators with instructions in program 2500 one or more pages at a time. For example, the instructions in page 2502 may be associated with performance indicators by scanning daemon 2508. Program 2500 is then executed by processor 2510. Data from the execution of program 2500 may then be collected. This data includes, for example, counts of events occurring in response to instructions in page 2502, counting the number of times each instruction in page 2502 is executed, and/or identifying the number of visits to page 2502.
Next, scanning daemon may remove the performance indicators from instructions in page 2502 and associate performance indicators with instructions in page 2504. Program 2500 is then executed again by processor 2510, and data from execution of this program is collected. Then, instructions in page 2506 may be modified in program 2500 executed to collect data on that page.
In this manner, usages of routines typically not recorded by programs, such as a timer profiler, may be identified. A timer profiler may not record some usages of routines because interrupts may be inhibited or the timing of samples may cause synchronous non-random behavior. By modifying instructions in program 2500, counting a routine or other modules may be obtained in which the counts are unbiased and the system is unperturbed. In this manner, interrupt driven counting is avoided. Further, although the instrumenting of code is one page at a time, other groupings of instructions may be used in scanning a program, such as modules that form the program. For example, the grouping may be a single executable program, a library, a group of selected functions, and a group of selected pages.
Turning next to
First, a selection of pages is identified (step 2600). In this example, the pages are those in the program that are to be scanned or instrumented. Next, a page within the selection of pages is selected for modification (step 2602). Indicators are then associated with all of the instructions in the selected page (step 2604). The program is then executed (step 2606). Next, a determination is made as to whether all the pages with the selection have been scanned (step 2608). If all of the pages have been scanned, the process terminates thereafter. However, if not all pages have been scanned, the next page to be scanned is selected (step 2610), with the process returning to step 2604 as described above.
The process illustrated in
A program is employed to identify a caller from a routine from the information found in a call stack. This program allows for an identification of what has occurred in a routine and provides a summary of what has occurred in a program by identifying function calls that have been made. This program, however, requires instructions inserted in the code to obtain this information.
The mechanism of the present invention allows for identifying calls and returns without having to perform special code instrumentation. In particular, the function of generating an interrupt on a specific set of instructions may be used to gather information about the system and applications. In these examples, instructions for calls and returns are associated with a performance indicator that generates an interrupt.
By walking back up the call stack, a complete call stack can be obtained for analysis. A “stack walk” may also be described as a “stack unwind”, and the process of “walking the stack” may also be described as “unwinding the stack.” Each of these terms illustrates a different metaphor for the process. The process can be described as “walking” as the process must obtain and process the stack frames step-by-step or frame-by-frame. The process may also be described as “unwinding” as the process must obtain and process the stack frames that point to one another, and these pointers and their information must be “unwound” through many pointer dereferences.
The stack unwind follows the sequence of function/method calls at the time of an interrupt and is generated in response to execution of an instruction associated with a performance indicator. A call stack is an ordered list of routines plus offsets within routines (i.e. modules, functions, methods, etc.) that have been entered during execution of a program. For example, if routine A calls routine B, and then routine B calls routine C, while the processor is executing instructions in routine C, the call stack is ABC. When control returns from routine C back to routine B, the call stack is AB. For more compact presentation and ease of interpretation within a generated report, the names of the routines are presented without any information about offsets. Offsets could be used for more detailed analysis of the execution of a program, however, offsets are not considered further herein.
Thus, during interrupt processing or at post-processing initiated by execution of an instruction associated with a particular performance indicator, the generated sample-based profile information reflects a sampling of call stacks, not just leaves of the possible call stacks, as in some program counter sampling techniques. A leaf is a node at the end of a branch, i.e. a node that has no descendants. A descendant is a child of a parent node, and a leaf is a node that has no children.
With reference now to
Call stack 2700 includes information identifying the routine that is currently running, the routine that invoked it, and so on, all the way up to the main program. Call stack 2700 includes a number of stack frames 2702, 2704, 2706, and 2708. In the depicted example, stack frame 2702 is at the top of call stack 2700, while stack frame 2708 is located at the bottom of call stack 2700. The top of the call stack is also referred to as the “root”. The interrupt (found in most operating systems) is modified to obtain the program counter value (pcv) of the interrupted thread, together with the pointer to the currently active stack frame for that thread. In the Intel architecture, this is typically represented by the contents of registers: EIP (program counter) and EBP (pointer to stack frame).
By accessing the currently active stack frame, it is possible to take advantage of the (typical) stack frame linkage convention in order to chain all of the frames together. Part of the standard linkage convention also dictates that the function return address be placed just above the invoked-function's stack frame; this can be used to ascertain the address for the invoked function. While this discussion employs an Intel-based architecture, this example is not a restriction. Most architectures employ linkage conventions that can be similarly navigated by a modified profiling interrupt handler.
When an interrupt occurs, the first parameter acquired is the program counter value. The next value is the pointer to the top of the current stack frame for the interrupted thread. In the depicted example, this value would point to EBP 2708a in stack frame 2708. In turn, EBP 2708 points to EBP 2706a in stack frame 2706, which in turn points to EBP 2704a in stack frame 2704. In turn, this EBP points to EBP 2702a in stack frame 2702. Within stack frames 2702-2708 are EIPs 2702b-2708b, which identify the calling routine's return address. The routines may be identified from these addresses. Thus, routines are defined by collecting all of the return addresses by walking up or backwards through the stack.
Obtaining a complete call stack may be difficult in some circumstances, because the environment may make tracing difficult, such as when an application having one call stack makes a call to a kernel having a different call stack. The hardware support provided by the mechanism of the present invention avoids some of these problems.
Turning next to
The process begins by identifying call and return instructions (step 2800). The instructions for calls and returns are ones of interest for determining when a routine has been called and when a routine completes. This may be accomplished for interrupts, interrupt returns, system calls, and returns from system calls.
Next, performance indicators are associated with the identified call and return instructions (step 2802). The program is then executed (step 2804), and data is collected from the performance monitor unit (step 2806) with the process terminating thereafter. This information may be collected through interfaces, such as hardware interface 2212 illustrated in
With this data, identifications of callers of routines may be made. This information may be used to generate data structures, such as trees to track and present information regarding the execution of the program. This generation of data structures may be implemented using processes similar to those provided in analysis tools.
Turning next to
First, a determination is made as to whether an execution of a selected instruction is detected (step 2900). This determination is made by examining each instruction that is executed to see whether a performance indicator is associated with the instruction. These performance indicators may be associated with the instructions through different tools, such as compiler 1500 in
If execution of an instruction containing a performance indicator is not identified, the process returns to step 2900 until a selected instruction is detected. If a selected instruction is identified as being executed, a counter with a set threshold is incremented for that selected instruction to count how often that particular instruction is executed (step 2902). In these examples, each instruction identified for monitoring is assigned a counter.
Next, a determination is made as to whether the set threshold has been reached (step 2904). Threshold values are initially determined by using documented cache miss times, for each of the cache levels. However, increasing times are used to determine problems caused by cache interventions (accesses from other processors). Repeated runs with different values may be made to identify the areas with the worst performance.
In these examples, the instruction may be associated with an indicator that includes an indication that execution of the instruction is to be monitored as well as providing a counter. Further, count criteria may be included to identify when an interrupt is to be generated. For example, an interrupt may be generated when the instruction has been executed more than thirteen times.
If the threshold has not been reached, the process returns to step 2900 as described above. If the set threshold has been reached, an interrupt is sent to the monitoring program (step 2906) with the process terminating thereafter. This interrupt may be sent to an interrupt unit, such as interrupt unit 250 in
This process may be especially useful for routines with many branches. In this case, all branch instructions would be flagged for counting. Information derived by this type of counting may be useful for identifying improvements for compiler and just-in-time (JIT) code generation by minimizing branches or adjusting hint flags, supported in the instruction architecture of the processor that is used.
Turning next to
First, a call stack is examined and the caller of a routine is identified (step 3000). Next, a count of the number of instructions executed is captured from the instruction cache (step 3002). The count is for a counter used in step 2902 in
Turning next to
Instruction cache 3106 uses range registers 3108 to define instruction ranges. These registers may be existing registers or instruction cache 3106 may be modified to include registers to define instruction ranges. These ranges may be based on addresses of instructions. Additionally, range registers 3108 may be updated by various debugger programs and performance tools.
If an instruction is executed in a range, such as instruction range 3102 or instruction range 3104, a counter is incremented in instruction cache 3106. Alternatively, the instruction may be sent to a performance monitor unit, such as performance monitor unit 240 in
Data accesses may be monitored in a similar fashion. For example, data 3112 includes data range 3114. Data accesses to data range 3114 may be counted in a similar fashion to execution of instructions within instruction range 3102 or instruction range 3104. These ranges may be defined in registers within a data unit, such as data cache 216 in
Turning next to
First, an instruction is identified for execution (step 3200). Next, a determination is made as to whether the instruction is within a set range of instructions (step 3202). The range may be identified by examining registers defining one or more instruction ranges. If the instruction is not within a set range of instructions, the process returns to step 3200 as described above. If the instruction is within a set range of instructions, a determination is made as to whether the previous instruction was within the set range of instructions (step 3204). If the previous instruction was not within the set range of instructions, a visit counter is incremented to tell the processor how many times the instruction range is entered (step 3206). Additionally, an execution counter is incremented to count the number of instructions executed within the set range of instructions (step 3208) with the process returning to step 3200 thereafter.
With reference again to step 3204, if the previous instruction was within the set range of instructions, the process proceeds to step 3208 as described above.
A similar process to the one illustrated in
The previous figures have primarily illustrated various ways in which performance indicators may be implemented to enable the counting of events within a processor when instructions that are associated with the performance indicators are executed or when data accesses that are associated with performance indicators are made. It would be particularly advantageous to employ this real-time performance information at the application-level because it is often difficult to acquire accurate production runtime models for the workloads that are experienced by a particular piece of software. Typically, a software developer tunes an application for an expected workload and then reconfigures the application as its computational environment changes.
For example, a software module may contain a cache. A common tuning exercise is to determine the amount of memory to allocate to the cache. Making the cache too large may take memory from other data structures, thereby inhibiting their ability to perform their tasks. On the other hand, making the cache too small may render the cache ineffective, thereby degrading overall performance by imposing cache maintenance overhead and then evicting cached elements before they are used due to space limitations. Many applications are quite sensitive to their cache efficiency because a cache hit is many times faster than a cache miss. Using the performance indicators as described above, an application can obtain information that enables it to monitor its own state and then dynamically adjust its behavior in real-time based upon runtime experience. Because the overhead of the performance monitor counters is so low, an application would be able to take advantage of a full spectrum of tracing and sampling tools, which themselves may be adjusted by the application as execution progresses.
The remaining figures primarily illustrate various ways in which these performance indicators may be used to dynamically gather performance information from the hardware such that the performance information is then available during runtime at the software level. Moreover, the performance information may be made available to the application whose performance is being monitored. In this manner, the software can autonomically change its behavior, particularly to enhance its performance.
With reference now to
With reference now to
Object code block 3404 contains instructions that are associated with performance indicators and/or instructions that access data areas that are associated with performance indicators. The execution of the instructions in object code block 3404 performs one or more actions that correspond to step 3302 in
Object code block 3406 contains instructions that obtain or read values from performance monitor registers that have been configured to count events in accordance with the performance indicators that are associated with object code block 3404 or related marked data. The execution of the instructions in object code block 3406 performs one or more actions that correspond to step 3304 in
Object code block 3408 contains instructions that determine whether the event counts that have been retrieved from the performance monitor satisfy logical conditions. The execution of the instructions in object code block 3408 performs one or more actions that correspond to step 3306 in
After the conditional expressions are evaluated, object code block 3410 contains instructions for executional branching to other sets of instructions based on the outcomes of those evaluations. For example, object code block 3412 contains instructions that are to be executed when a first condition is satisfied; the execution of the instructions in object code block 3412 performs one or more actions that correspond to step 3308 in
With reference now to
Statement 3502 declares a definition of a subroutine. Statement 3504 obtains data from an input datastream and places the data items from the datastream into a data structure while counting the number of data items that have been inputted. At some point in time, object code for statement 3504 would have been generated to include instructions, wherein those instructions (or the data areas that are accessed by those instructions) have been associated with performance indicators; the performance indicators may have been generated statically by a compiler, e.g., as described above with respect to
Statement 3506 calls a function within an API to read values from two counters within the performance monitor. It may be assumed that variable names 3508 and 3510 are constant values that are resolved by the compiler to values or appropriate indicators that identify the counters/registers within the performance monitor from which the requested values are to be retrieved; the retrieved values are then returned to the calling subroutine via the other subroutine arguments, i.e., variables 3512 and 3514.
Statement 3516 is an if-then type of statement that evaluates a logical expression and then causes the execution control to pass through the statement or to flow around the statement, depending upon whether the outcome of the evaluation is a logical true value or a logical false value. In this example, if the logical expression is true, i.e., the counted number of cache misses is greater than the counted number of cache hits, then statement 3518 calls a subroutine that will modify the data structure. Constant value 3520 indicates that the subroutine is merely to adjust or to tweak the data structure in some programmatic manner to make the use of the data structure slightly more efficient in an effort to reduce the number of cache misses that are generated while using the data structure, which would depend upon the particular data structure and other factors in the computational environment.
Statement 3522 is a continuation of the if-then type of statement. If the logical expression is true, which in this case means that the counted number of cache misses is greater than some number that is the product of a constant and the number of data items that were stored into the data structure, then statement 3524 calls a subroutine that will modify the data structure. Constant value 3526 indicates that the subroutine should rebuild the data structure in some programmatic manner. It may be assumed that the subroutine acts to modify the configuration of the data structure in which the newly added data items were stored in order to greatly reduce the number of cache misses. After the data structure has been modified in some manner, statement 3528 then calls a subroutine that will sort the new data items in the data structure in some manner.
In this example, the data structure might be a hybrid of two different types of common data structures. Each of the two parts of the data structure may be used to obtain the advantages of their different characteristics. The data structure may be frequently searched for information about the data items, but the data items may be removed at any time. For example, data items that have been newly added might be much more likely to be removed; hence, it may not be efficient to store the newly added data items in a sorted, semi-permanent manner with the other data items that have been present within the data structure for a relatively long period of time. In order to handle the data items in an efficient manner, one part of the data structure may store data items that have been present within the data structure for a relatively long period of time, and another part of the data structure may store data items that have been newly added to the data structure and might be deleted relatively quickly. The call to the subroutine in statement 3524 to rebuild the data structure may be an operation that restructures the part of the data structure that contains the newly added data items.
The example in
However, the selected characteristics might not be substantially related to the efficiency of the use of the data structure. If the chosen threshold is too small, then the data structure might be rebuilt too often, yet if the chosen threshold is too high, the data structure might not be rebuilt often enough. In other words, for this second case, if too many data items are added to the data structure before it is rebuilt or restructured, then the data structure might grow so large that the cache is not used efficiently, thereby incurring many cache misses when the data structure is used.
In contrast, using the embodiments of the present invention that are illustrated in
In the description of
Referring again to the pseudocode in
Compiler directive 3530 is merely an example of a possible format for allowing a software developer to embed information into the source code of an application such that a compiler generates the appropriate object code that includes the performance indicators that are being requested by the software developer. In the example that is shown in
The performance monitor would continue to count events until it was configured to do otherwise, e.g., by the processing of performance indicators that are associated with instructions that are executed at some later point in time that configure the performance monitor to stop counting those events. In other words, in the example that is shown in
With reference now to
For the above-noted embodiment, compiler directives are necessary because the source code language does not include support for performance information gathering operations; compiler directives are external to the source code language, yet compiler directives assist a compiler in generating object code from source code statements. In an alternative embodiment, the source code language may be extended to provide support for performance information gathering operations, e.g., by including reserved words similar to other reserved words in the source code language.
As mentioned above, a compiler generates object code that is intended to be executed on a specific class of CPU, notwithstanding the generation of bytecode by certain types of compilers, wherein the bytecode is intended to be interpreted within a standardized virtual machine such that the bytecode is eventually runnable on many different classes of CPUs. When a compiler generates object code, the compiler may have several choices for generating certain aspects of the object code; different compilers will generate different object code modules from an identical source code file because they make different choices in the manner in which certain aspects of the software are implemented. For example, a compiler can allocate data areas for different types of variables in different memory locations; the compiler can also determine whether those data areas should be allocated statically at compile-time or dynamically at runtime. The choice of memory locations and their relative arrangement will influence the performance of the instructions that access those data areas, e.g., creating fewer or more cache hits to the CPU's on-chip cache. Similarly, groups of instructions are implicitly generated for certain types of source code logic, e.g., iterative loops, and the compiler has the ability to place groups of instructions in different areas through the use of branch instructions or jump instructions. Assuming that a particular CPU has an on-chip instruction cache, the compiler can influence the performance of groups of instructions based on their relative location; for example, two groups of instructions that are juxtaposed in memory and executed sequentially might generate zero instruction cache misses if both groups of instructions will easily fit into the instruction cache.
Although a compiler generates object code for a particular class of CPUs, the capabilities of a particular CPU, though, may vary widely within a class of CPUs that adhere to a similar architecture. For example, a particular CPU may have a larger on-chip data cache than a similar CPU that executes the same instruction set at the same processor clock speed. At compile-time, the compiler does generally not have any information about the capabilities or configuration of a particular CPU that will eventually execute the object code that is being generated by the compiler, so the compiler generates object code with certain characteristics that seem appropriate for any instance within that class of CPUs.
However, using the present invention, a compiler can shift compile-time decisions to runtime decisions. Rather than statically determining certain characteristics of the object code when the object code is generated, which thereby also inherently determines the performance of the execution of the object code to some degree, certain characteristics of the object code can be determined when the object code is executed by allowing the application to use knowledge about itself to affect its behavior. The compiler configures the software application to obtain real-time information about the performance of the software application within the low-level hardware, and the compiler configures the software application to dynamically change its behavior based on this real-time performance information, which is illustrated with respect to
With reference now to
At some point during the generation of object code for the subroutine, the compiler determines that it may generate the object code in different ways or forms, possible using different techniques (step 3704); e.g., as mentioned above, the compiler might determine that it has multiple options for selection locations of data objects, locations of groups of instructions, manners of branching or jumping between groups of instructions, etc. Rather than selecting one choice among each option that occurs, the compiler generates multiple alternative versions of object code for the subroutine (step 3706). A software developer may control the ability of the compiler to perform such operations through special compiler directives. The software developer might insert a compiler directive within the source code of a subroutine that explicitly allows the compiler to generate multiple alternative versions of object code for the subroutine. Similarly, the software developer might insert a compiler directive within the source code of a subroutine that explicitly disallows the compiler from generating multiple alternative versions of object code for the subroutine. As another alternative, the software developer might be required to place a special compiler directive at the start of a source code file in order to explicitly permit the compiler to generate multiple alternative versions of object code for any subroutine within the source code file.
The compiler manages the multiple alternative versions of object code for the subroutine by assigning different identifiers to the different versions of the subroutine (step 3708). In other words, the compiler can manage each version of the subroutine's object code as if the object code had originated from different subroutines. However, the compiler must adjust the remainder of the software application to account for the multiple versions; for each call to the subroutine within the software application, the compiler must ensure that a call is made to one of the multiple alternative versions of object code for the subroutine.
The manner in which the compiler adjusts the original calls to the subroutine is linked to the differences that the compiler has introduced to the multiple alternative versions of object code for the subroutine. The compiler has generated multiple versions of object code for the subroutine because the compiler has determined that each version may have certain performance-enhancing characteristics; e.g., as noted above, the locations of instructions and data objects may reduce the number of instruction cache misses or data cache misses. Hence, the compiler replaces each call to the original subroutine with a call to one of the multiple alternative versions based upon a dynamically determined characteristic that is appropriate for the call that is being replaced or for the version of the subroutine that is to be selected to be called. More specifically, the compiler replaces each call to the subroutine with a selective or condition call to one of the multiple alternative versions based on information gathered from performance indicators (step 3710). In other words, during the execution of the object code, event counts would be gathered through the use of performance indicators just prior to a call to one of the multiple versions of object code for the subroutine; depending on those event counts, a determination is made as to which subroutine should be called, as illustrated in more detail further below.
With reference now to
With reference now to
With reference now to
The first object code subsection 4004 within object code section 4002 contains instructions that are associated with performance indicators and/or instructions that access data associated with performance indicators; these performance indicators configure the performance monitor to begin counting hardware events associated with the execution of those instructions. The second object code subsection 4006 within object code section 4002 contains instructions that obtain or read the performance monitor counters that have been counting the hardware events. The third object code subsection 4008 within object code section 4002 contains instructions that compute conditional expressions based on the values from the performance monitor counters. The fourth object code subsection 4010 within object code section 4002 contains instructions for calling or jumping to an appropriate version of the given subroutine based on the computed conditional expressions.
With reference to
Compiler directive 4102 causes performance indicators to be associated with data accesses to a particular array, which may be implemented using marked data in a variety of manners as previously described above. In this example, compiler directive 4102 indicates that particular hardware events, i.e., data cache misses, are to be counted when accessing the identified array, e.g., each time an instruction accesses a memory location in a range of memory addresses that is reserved for the array. Statement 4104 reads the last element from the identified array, thereby potentially causing a cache miss event to occur. In this example, the compiler has generated object code that contains a simple test to check whether accessing an element of the array generates a cache miss event.
As mentioned previously, the compiler replaces each call to the original subroutine with a call to one of the multiple alternative versions based upon a dynamically determined characteristic that is appropriate for the call that is being replaced or for the version of the subroutine that is to be selected during runtime to be called. In the example that is shown in
As mentioned above, different sets of the remaining figures illustrate different uses at the software level for information gathered at the hardware level while employing performance indicators. Turning now to
An operating system allocates time slices to threads in accordance with their needs and their competition for shared resources, and an operating system typically provides multiple mechanisms for coordinating the use of shared resources by multiple threads. A common mechanism for serializing access to a shared resource is a mutex, or mutual exclusion lock, which is a simple lock having two states: locked and unlocked. The lock is typically implemented as a simple data value that is stored in memory. A mutex can be logically associated with a shared resource such that a thread that successfully locks the mutex is said to be the current owner of the mutex; only the thread that possesses a particular mutex should proceed to access the shared resource that is associated with that particular mutex, and only the thread that possesses a particular mutex should unlock that particular mutex. Thus, a critical section of code within a thread that accesses a shared resource is bounded by a call to lock a mutex and a call to unlock the same mutex. If a thread attempts to lock a mutex and fails, then it must wait until it is able to lock the mutex before proceeding to execute its critical section of code in which it accesses the shared resource. A mutex can be used to synchronize threads within a single process or across multiple processes if the mutex is allocated within memory that is shared by the coordinating processes.
The manner in which a thread waits for a mutex after failing to acquire the mutex depends on the manner in which the mutex mechanism is implemented. Three types of locks are encountered: a blocking lock, a spin lock, and some type of combination of a blocking lock and a spin lock. If a mutex has already been acquired and another thread requests to lock the mutex, then a mutex that is implemented as a blocking lock causes the waiting thread to cease being executable or to be suspended, i.e., to go to “sleep”. In contrast, spin locks do not put waiting threads to sleep. Instead, a waiting thread executes a loop, thereby repeatedly requesting the lock until it is freed by the thread that currently owns the mutex; the loop may contain an empty, iterative loop, i.e., “busy loop” or “busy wait”, that increments or decrements a variable such that the thread does not immediately re-request the mutex but waits for a period of time that depends on the length of the iterative loop.
A semaphore is another locking mechanism for restricting access to shared resources, but in contrast to a mutex, a typical semaphore allows one or more threads to access a shared resource simultaneously. Even though a first thread may already be accessing a shared resource, additional threads are allowed to access the shared resource, up to a maximum number. A semaphore acts as a specialized counter, whereas a mutex acts as a toggle switch. Hence, a semaphore can be regarded as a more generalized locking mechanism, and a mutex is sometimes described as a mutex semaphore or a binary semaphore; hereinbelow, the present invention is operable with generalized semaphores, so it is assumed that a semaphore may act as a mutex.
With reference now to
With reference now to
Instruction 4300 copies the value of the semaphore from a memory address to the accumulator within the CPU; the memory address is provided by a variable name that can be resolved by a compiler or an assembler to a memory address. Instruction 4302 copies the value from the accumulator to a register, and instruction 4304 increments the value in the register.
Instruction 4306 is a type of compare-and-exchange instruction that requires a source operand and a destination operand. This instruction executes as follows. If the value in the accumulator equals the value at the identified memory address, then the value in the register (source operand) is copied to the identified memory address (destination operand), and the “zero” processor status flag is set; the memory address is again provided by a variable name that represents the semaphore. If the value in the accumulator does not equal the value at the identified memory address, then the value at the identified memory address is loaded into the accumulator, and the “zero” processor status flag is cleared; this case may occur if another thread has modified the semaphore between the execution of instruction 4300 and the execution of instruction 4306, e.g., the thread with instructions 4300-4306 was swapped out at some point in time, which allowed another thread to execute and change the value of the semaphore.
Instruction 4308 jumps to the identified address if the “zero” processor status flag is set, i.e., the program counter is changed to the identified address so that the next instruction is retrieved from that address. Instruction 4310 jumps to the identified address if the “zero” processor status flag is cleared, i.e., not set. If the thread has successfully acquired the semaphore, then instruction 4312 is executed, which represents the beginning of an execution path during which the thread will access a shared, restricted resource. Hence, instruction 4312 represents an instruction that is executed if the semaphore is successfully acquired, and instruction 4310 represents an instruction that is executed if the thread fails to acquire the semaphore.
If the semaphore has been successfully acquired, its value has been incremented. The value of the semaphore is protected from exceeded its maximum value by instructions 4314-4318. Instruction 4314 copies the value of the semaphore from the accumulator to a register, and instruction 4316 subtracts a constant value from the register; the constant value is identified by a name that can be resolved by a compiler or an assembler to a pre-determined value. Instruction 4318 jumps to the identified address if the “zero” processor status flag is set, which in this case indicates that the semaphore has reached is maximum value. The thread remains in a tight initial loop until the value of the semaphore is no longer at its maximum.
With reference now to
The scanning daemon then associates performance indicators with instructions that are related to acquiring the semaphores but at three different points for three different purposes: instructions that reserve or modify a semaphore (step 4404); instructions that are executed if a semaphore is acquired (step 4406); and instructions that are executed if a semaphore is not acquired (step 4408), thereby concluding the process. Although these examples show that the performance indicators are associated with instructions through the operations of a scanning daemon, in alternative embodiments, the performance indicators may be associated with instructions through different types of operations, as discussed further above.
With reference now to
With reference now to
In contrast to
Eventually, count values 4632, 4634, and 4636 can be obtained from performance monitor 4624 to represent the number of attempts to acquire the semaphore, the number of failed attempts, and the number of successful attempts, respectively. It is possible for an application to count the number of attempts, successes, and failures, but if a software entity other than the application wanted these counts, an interface would be required to request and transfer the counts. With the present invention, the counts may be gathered through the use of the performance indicators. These count values represent statistical information about the performance of the operations on the semaphore, and these count values may be used by various software entities, particularly for improving the performance of operations associated with the semaphore. Reports on semaphore utilization may be used by an analyst to change the algorithms related to the semaphores; for example, multiple semaphores might be deployed in place of a single semaphore to protect different portions of data in order to reduce contention. Moreover, the statistical information can be gathered in an efficient, non-intrusive manner such that the application does not need to explicitly generate these counts.
As mentioned above, different sets of the remaining figures illustrate different uses at the software level for information gathered at the hardware level while employing performance indicators. Turning now to
The performance of an application is affected by the spatial locality that is exhibited by the application during its execution. By conglomerating the application's hot code blocks and hot data areas together, there is a higher likelihood that those hot code blocks and hot data areas will be located in a cache when they are needed. In other words, aggregating hot spots improves the performance of an application by reducing the number of instruction cache misses, data cache misses, translation lookaside buffer (TLB) misses, etc.
There are many techniques in the prior art for enhancing application performance by improving spatial locality to achieve these types of characteristics, such as better utilization of hardware caches, instruction pipelines, etc. For example, the movement of application blocks in various ways is described in Pettis et al., U.S. Pat. No. 5,212,794, issued May 18, 1993, entitled “Method for Optimizing Computer Code to Provide More Efficient Execution on Computers Having Cache Memories”, hereby incorporated by reference, and the optimization of memory references in various ways is described in Heisch, U.S. Pat. No. 5,689,712, issued Nov. 18, 1997, entitled “Profile-Based Optimizing Postprocessors for Data References”, hereby incorporated by reference.
However, these prior art techniques do not allow the optimization to occur while the application is running. The prior art uses instrumentation and trace techniques to collect data about the application during runtime, and the collected data is analyzed during a post-processing phase; any optimizations are applied while the application is not running, and the enhanced performance in the application can only be realized during the next time that the application is run. In contrast, in the present invention, the memory manager may improve the spatial locality of an application during a runtime phase by collecting performance information during runtime, analyzing the information during runtime, and modifying the application's memory space during runtime, thereby dynamically providing enhanced performance.
With reference now to
The memory manager then generates and sends a performance monitoring request to a performance scanning daemon that is similar to that described in
At some later point in time, the memory manager receives a performance monitoring response from the scanning daemon (step 4708). The performance monitoring response may be transferred in a variety of ways that is appropriate for the interface between the memory manager and the scanning daemon. If the performance monitoring request was made by a subroutine call, then the memory manager may have registered, during the initial call, a callback subroutine that the scanning daemon may call to return the performance monitoring response. If the performance monitoring request was passed as a message, then the performance monitoring response may be passed as a message. It may be assumed that the request and the response are transmitted in an asynchronous manner such that the memory manager may perform other operations while waiting for the performance monitoring response.
After receiving the performance monitoring response, the memory manager examines the information within the response to determine hot spots within the monitored application (step 4710). While the performance monitoring request was being processed by the scanning daemon, the scanning daemon was collecting events counts for data accesses, instruction accesses, instruction executions, or other types of events as requested by the memory manager. In addition, these events have been counted over a particular period of time, so the information that is generated and returned to the memory manager reflect events per period of time, or event rates. By comparing event rates to other concurrent event rates, historical event rates, configurable threshold values, configurable conditional rules, etc., the memory manager is able to determine which code segments and data areas within an application are hot or cold, i.e., executed or accessed frequently or infrequently. After determining the hot spots, the memory manager conglomerates the hot code blocks and hot data areas (step 4714), thereby concluding the process. The memory manager may repeat this process for other applications or for the same application after a period of time.
With reference now to
The process commences when the scanning daemon receives a performance monitoring request from a memory manager (step 4802). Using the memory locations of object code blocks and data areas as indicated in the performance monitoring request, the scanning daemon associates performance indicators with those memory locations as requested (step 4804); the performance indicators may be stored in memory, shadow memory, cache, etc., as is appropriate for the hardware architecture that supports the present invention. In addition, as requested by the memory manager, the scanning daemon limits the amount of time during which the performance indicators allow for the counting of events, e.g., by setting software timers for a requested period of time (step 4806).
At the end of that time period, the scanning daemon then retrieves the counts that have been gathered for the events that were generated in accordance with the performance indicators (step 4808); for example, the scanning daemon reads the counts from registers within a performance monitor in the CPU. It may be assumed that the scanning daemon set the timers in an asynchronous manner such that the scanning daemon is able to perform other operations during that time period. In addition, the scanning daemon is able to handle multiple concurrent requests; the requests may initiate the monitoring of different locations within an application space or the monitoring of different application memory spaces. Alternatively, the scanning daemon may determine the length of the time period and then return that value to the requester as part of the response. The scanning daemon then returns the counts that have been gathered in a performance monitoring response to the memory manager (step 4810), and the process is concluded.
With reference now to
In this example, memory manager 4902 comprises multiple entities as a convenience for illustrating various functional units that may be present within a data processing system; memory manager may include fewer or more functional units, depending on the operating system, or the functionality may be organized in some other manner. Dynamic loader 4904 loads application modules into memory. Memory allocator 4906 allocates memory to an application upon request. Garbage collector 4908 gathers memory blocks that were previously allocated to an application but that are no longer being used.
In this embodiment of the present invention, memory relocator 4910 moves object code blocks and data blocks in order to improve the performance of the application in a manner similar to that described above with respect to
After some period of time, scanning daemon 4914 gathers the counts that represent the number of hardware events that occurred in accordance with the processing of the performance indicators that indicate the hardware events that should be counted. These counts, possibly along with other information, are placed into performance monitoring response 4930 by scanning daemon 4914 and returned to memory relocator 4910.
Memory relocator 4910 analyzes the execution rates or access rates of the previously chosen memory locations or memory areas to determine if any of the code blocks or data areas should be moved in an attempt to improve the performance of the application.
However, it is not necessary that the code blocks or data areas should moved as a whole. Depending on the granularity of the information that was gathered via the performance indicators, in some cases, the code blocks and data areas might be split, thereby allowing the conglomeration of hot areas and the conglomeration of cold areas. It should also be noted that the memory relocation operation may act in accordance with many different configurable conditions and rule sets. For example, it is possible that the memory relocator determines that many code blocks appear to execute with the same relative frequency with a few code blocks that are rarely executed; in that case, the memory relocation operation may be performed in order to move those relatively cold blocks away from the other code blocks rather than attempting to aggregate hot blocks.
Thus, the present invention provides an improved method, apparatus, and computer instructions for providing assistance in monitoring execution of programs. The mechanism of the present invention includes employing an indicator that is recognized by the processor to enable counting events that are associated with the execution of an instruction or events that are associated with accessing a particular memory location. Various types of counting as described above are enabled through this mechanism. Further, with the information provided through the use of associating indicators with particular instructions, the mechanism of the present invention also provides for various types of adjustments to programs in monitoring and analyzing performance of programs. Further, as described above, programs may be automatically adjusted to allow for monitoring of selected instructions and even routines and modules without having to modify the program.
It is important to note that while the present invention has been described in the context of a fully functioning data processing system, those of ordinary skill in the art will appreciate that the processes of the present invention are capable of being distributed in the form of a computer readable medium of instructions and a variety of forms and that the present invention applies equally regardless of the particular type of signal bearing media actually used to carry out the distribution. Examples of computer readable media include recordable-type media, such as a floppy disk, a hard disk drive, a RAM, CD-ROMs, DVD-ROMs, and transmission-type media, such as digital and analog communications links, wired or wireless communications links using transmission forms, such as, for example, radio frequency and light wave transmissions. The computer readable media may take the form of coded formats that are decoded for actual use in a particular data processing system.
The description of the present invention has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art. For example, instead of using a field in an instruction or in a bundle, a new instruction or operation code may be used to indicate that a subsequent instruction, or a subsequent set of instructions are marked instructions. Also, the architecture of a processor may be changed to include additional bits if spare fields for performance indicators are unavailable in the case in which it is desirable to include performance indicators within fields in the instructions. Also, although examples of events, such as execution of the instruction, time, such as clock or processor cycles, needed to execute an instruction, time to access data, entry into a section of code, have been given, these examples are not meant to limit the present invention to the types of events that can be counted. Any event relating to execution of an instruction or access to a memory location may be counted using the mechanisms of the present invention.
The illustrative embodiments were chosen and described in order to best explain the principles of the invention, the practical application, and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated.
The present invention is related to the following applications: “Method and Apparatus for Selectively Counting Instructions and Data Accesses”, Ser. No. ______, attorney docket no. AUS920030478US1; “Method and Apparatus for Generating Interrupts Upon Execution of Marked Instructions and Upon Access to Marked Memory Locations”, Ser. No. ______, attorney docket no. AUS920030479US1; “Method and Apparatus for Counting Data Accesses and Instruction Executions that Exceed a Threshold”, Ser. No. ______, attorney docket no. AUS920030480US1; “Method and Apparatus for Counting Execution of Specific Instructions and Accesses to Specific Data Locations”, Ser. No. ______, attorney docket no. AUS920030481US1; “Method and Apparatus for Debug Support for Individual Instructions and Memory Locations”, Ser. No. ______, attorney docket no. AUS920030482US1; “Method and Apparatus to Autonomically Select Instructions for Selective Counting”, Ser. No. ______, attorney docket no. AUS920030483US1; “Method and Apparatus to Autonomically Count Instruction Execution for Applications”, Ser. No. ______, attorney docket no. AUS920030484US1; “Method and Apparatus to Autonomically Take an Exception on Specified Instructions”, Ser. No. ______, attorney docket no. AUS920030485US1; “Method and Apparatus to Autonomically Profile Applications”, Ser. No. ______, attorney docket no. AUS920030486US1; “Method and Apparatus for Counting Instruction and Memory Location Ranges”, Ser. No. ______, attorney docket no. AUS920030487US1; “Method and System for Autonomic Execution Path Selection in an Application”, Ser. No. ______, attorney docket no. AUS920030489US1; “Method and System for Autonomic Performance Improvements in an Application Via Memory Relocation”, Ser. No. ______, attorney docket no. AUS920030490US1; all of which are assigned to the same assignee, and incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10682437 | Oct 2003 | US |
| Child | 12137430 | US |
