Patent Application
20100318703
The invention relates to the field of computers and telecommunications, and more particularly to an advanced processor for use in computers and telecommunications applications.
Modern computers and telecommunications systems provide goat benefits including the ability to communicate in form at ion around the world. Conventional architectures for computers and telecommunications equipment include a large number of discrete circuits, which causes inefficiencies in both the processing capabilities and the communication speed.
For example, HG. 1 illustrates such a conventional line card employing a number of discrete chips and technologies. In
Advances in processors and other components have improved the ability of telecommunications equipment to process, manipulate, store, retrieve and deliver information. Recently, engineers have begun to combine functions into integrated circuits to reduce the overall number of discrete integrated circuits, while still performing the required functions at equal or bettor levels of performance. This combination has been spurred by the ability to increase the number of transistors on a chip with new technology and the desire to reduce costs. Some of these combined integrated circuits have become so highly functional that they are often referred to as a System on a Chip (SoC). However, combining circuits and systems on a chip can become very complex and pose a number of engineering challenges. For example, hardware engineers want to ensure flexibility for future design and software engineers want to ensure that their software will run on the chip and future designs as well.
The demand for sophisticated new networking and communications applications continues to grow in advanced switching and routing. In addition, solutions such as content-aware networking, highly integrated security, and new forms of storage management are beginning to migrate into flexible multi-service systems. Enabling technologies for these and other next generation solutions must provide intelligence and high performance with the flexibility for rapid adaptation to new protocols and services.
In order to take advantage of such high performance networking and data processing capability, it is important that such systems be capable of communicating with a variety of high bandwidth peripheral devices, preferably using a standardized, high-bandwidth bus. Although many proprietary high-bandwidth buses are possible, using a standardized bus allows the system to interface with a broader variety of peripherals, and thus enhances the overall value and utility of the system.
One high-bandwidth standardized bus that has become popular in recent years is the PCI Express (PCI-E or PCIe) interface. The PCIe interface, originally proposed by Intel as a replacement for the very popular but bandwidth limited personal computer PCI interface, is both high bandwidth and, due to the fact that it has now become a standard component on personal computer motherboards, very widely adopted. Hundreds or thousands of different peripherals are now available that work with the PCIe interface, making this interface particularly useful for the present advanced processing system.
In contrast to the earlier, parallel, PCI system, which encountered bandwidth limitations due to problems with keeping the large number of parallel circuit lines in synchronization with each other at high clock speeds, the PCIe system is a very fast serial system. Serial systems use only a very small limited number of circuit lines, typically two to transmit and two to receive, and this simpler scheme holds up better at high clock speeds and high data rates. PCIe further increases bandwidth by allowing for multiple serial circuits. Depending upon the PCIe configuration, there can be as few as 1 bidirectional circuit, or as many as 32 bidirectional serial circuits.
Although, on a hardware level, the serial Me system is radically different from the earlier parallel PCI system, the earlier PCI system was extremely successful, and the computer industry had made a massive investment in earlier generation PCI hardware and software. To help make the much higher bandwidth PCIe system compatible with the preexisting PCI hardware and software infrastructure, PCIe was designed to mimic much of the earlier parallel PCI data transport conventions. Earlier generation software thus can continue to address PCIe devices as if they were PO devices, and the PCIe circuitry transforms the PCI data send and receive requests into serial PCIe data packets, transmits or receives these data packets, and then reassembles the serial PCIe data packets back into a format that can be processed by software (and hardware) originally designed for the PCI format. The PCIe design intention of maintaining backward compatibility, while providing much higher bandwidth, has been successful and PCIe has now become a widely used comp titer industry standard.
Although other workers, such as Stufflebeam (U.S. Pat. No. 7,058,738) have looked at certain issues regarding interfacing multiple CPUs to multiple I/O devices through a single switch (such as a PCIe switch), this previous work has focused on less complex and typically lower-performance multiple CPU configurations, that do not have to contend with the issues that result when multiple cores must coordinate their activity via other high-speed (and often on-chip) communication ring and interconnects.
Consequently, what is needed is an advanced processor that can take advantage of the new technologies while also providing high performance functionality. Additionally, this technology would be especially helpful it included flexible modification ability, such as the ability to interface with multiple high-bandwidth peripheral devices, using high-bandwidth star topology buses such as the PCIe bus.
The present invention provides useful novel structures and techniques for overcoming the identified limitations, and provides an advanced processor that can take advantage of new technologies while also providing high performance functionality with flexible modification ability. The invention employs an advanced architecture System on a Chip (SoC) including modular components and communication ion Structures to provide a high performance device.
This advanced processor comprises a plurality of multithreaded processor cores each having a data cache and instruction cache. A data switch interconnect (DSI) is coupled to each of the processor cores by its respective data cache, and configured to pass information among the processor cores. A level 2 (L2) cache, a memory bridge, and/or a super memory bridge, can also be coupled to the data switch interconnect (DSI) and configured to store information accessible to the processor cores.
A messaging network is coupled to each of the processor cores by the cores respective instruction caches (message station). A plurality of communication ports are connected to the messaging network. In one aspect of the invention, the advanced telecommunications processor further comprises an interface switch interconnect (ISI) coupled to the messaging network and the plurality of communication ports and configured to pass information among the messaging network and the communication ports. This interface switch interconnect may also communicate with the super memory bridge. The super memory bridge may also communicate with one or more communication ports and the previously discussed DSI.
In the embodiment of the invention disclosed here, the messaging network and the connect to a PCI express (PCIe) interface, enabling the processor to interface with a broad variety of high bandwidth PCIe peripherals.
Advantages of the PCIe embodiment of the invention include the ability to provide high bandwidth communications between computer systems and a large number of peripherals in an efficient, flexible, and cost-effective manner.
The invention is described with reference to the FIGS, in which:
The invention is described with reference to specific architectures and protocols. Those skilled in the art will recognize that the description is for illustration and to provide the best mode of practicing the invention. The description is not meant to be limiting and references to telecommunications and other app heat ions may be equally applicable to general computer applications, for example, server applications, distributed shared memory applications and so on. As described herein, reference is made to Ethernet Protocol, Internet Protocol, Hyper Transport Protocol and other protocols, but the invention may be applicable to other protocols as well. Moreover, reference is made to chips that contain integrated circuits while other hybrid or meta-circuits combining those described in chip form is anticipated. Additionally, reference is made to an exemplary MIPS architecture and instruction set, but other architectures and instruction sets can be used in the invention. Other architectures and instruction sets include, for example, x86, PowerPC, ARM and others,
The architecture is focused on a system designed to consolidate a number of the functions performed on the conventional line card and to enhance the line card functionality. In one embodiment, the invention is an integrated circuit that includes circuitry for performing many discrete functions. The integrated circuit design is tailored for communication processing. Accordingly, the processor design emphasizes memory intensive operations rather than computationally intensive operations. The processor design includes an internal network configured for relieving the processor of burdensome memory access processes, which are delegated to other entities for separate processing. The result is a high efficient memory access and threaded processing.
Again, one embodiment of the invention is designed to consolidate a number of the functions performed on the conventional line card of
The exemplary processor is designed as a network on a chip. This distributed processing architecture allows components to communicate with one another and not necessarily share a common clock rate. For example, one processor component could be clocked at a relatively high rate while another processor component is clocked at a relatively low rate. The network architecture further supports the ability to add other components in future designs by simply adding the component to the network. For example, if a future communication interface is desired, that interface can be laid out on the processor chip and coupled to the processor network. Then, future processors can be fabricated with the new communication interface.
The design philosophy is to create a processor that can be programmed using general purpose software tools and reusable components. Several exemplary features that support this design philosophy include: static gate design; low-risk custom memory design; flip-flop based design; design-for-testability including a full scan, memory built-in self-test (BIST), architecture redundancy and tester support features; reduced power consumption including clock gating; logic gating and memory banking datapath and control separation including intelligently guided placement; and rapid feedback of physical implementation.
The software philosophy is to enable utilization of industry standard development tools and environment. The desire is to program the processing using general purpose software tools and reusable components. The industry standard tools and environment include familiar tools, such as gcc/gdb and the ability to develop in an environment chosen by the customer or programmer.
The desire is also to protect existing and future code investment by providing a hardware abstraction layer (HAL) definition. This enables relatively easy porting of existing applications and code compatibility with future chip generations.
Turning to the CPU core, the core is designed to be MIPS64 compliant and have a frequency tart in the range of approximately 1.5 GHz+. Additional exemplary features supporting the architecture include: 4-way multithreaded single issue 10-stage pipeline; real time processing support including cache line locking and vectored interrupt support; 32 KB 4-way set associative instruction cache; 32 KB 4-way set associative data cache; and 128-entry translation-lookaside buffer (TLB).
One of the important aspects of the exemplary embodiment is the high-speed processor input/output (I/O), which is supported by: two XGMII/SPI-4 (e.g., boxes 228a and 228b of
Use of high bandwidth star topology serial communication buses, such as the PCIe bus, is useful because it helps expand the power and utility of the processor. Before discussing how PCIe technology may be integrated into this type of processor, a more detailed review of PCIe technology is warranted.
As previously discussed, the PCIe bus is composed of one or more (usually multiple) high-bandwidth bidirectional serial connections. Each bidirectional serial connection is called a “lane”. These serial connections are in turn controlled by a PCIe switch that can create multiple point-to-point serial connections between various PCIe peripherals (devices) and the PCIe switch in a star topology configuration. As a result, each device gets its own direct serial connection to the switch, and doesn't have to share this connection with other devices. This topology, along with the higher inherent speed of the serial connection, helps increase the bandwidth of the PCIe bus. To further increase bandwidth, a PCIe device can connect to the switch with between 1 and 32 lanes. Thus a PCIe device that needs higher bandwidth can utilize more PCIe lanes, and a PCIe device that needs lower bandwidth can utilize fewer PCIe lanes.
Note that the present teaching of efficient means to interface serial buses to this type of advanced processor is not limited to the PCIe protocol per-se. As will be discussed, these techniques and methods can be adapted to work with a broad variety of different star topology high-speed serial bus protocols, including HyperTransport, InfiniBand, RapidIO, and StarFabric. To simplify discussion, however, PCIe will be used throughout this disclosure as a specific example.
As previously discussed, in order to maximize backward compatibility with the preexisting PCI parallel bus, PCIe designers elected to make the new serial aspects of the PCIe bus as transparent (unnoticeable) to system hardware and software as possible. They did this by making the high level interface of the PCIe bus resemble the earlier PCI bus, and put the serial data management functions at a lower level, so that preexisting hardware and software would not have to contend with the very different PCIe packet based serial data exchange format. Thus PCIe data packets send a wide variety of different signals, including control signals and interrupt signals.
Detailed information on the PCI express can be found in the book “PCI express system architecture” by Budruk, Anderson, and Shanley, Mindeshare, Inc. (2004), Addison Wesley.
Briefly, the PCIe protocol consists of the Physical layer (containing a logical sublayer and an electrical sublayer), the transaction layer, and a data link layer.
The physical layer (sometimes called the PIPE or PHY) controls the actual serial lines connecting various PCIe devices. It allows devices to form one or more serial bidirectional “lanes” with the central PCIe switch, and utilizes each device's specific hardware and bandwidth needs to determine exactly how many lanes should be allocated to the device. Since all communication is by the serial links, other messages, such as interrupts and control signals are also sent by serial data packets through these lanes. Instead of using clock pulses to synchronize data (which use a lot of bandwidth), the data packets are instead sent using a sequential 8-bit/10-bit encoding scheme which in itself carries enough clock information to ensure that the devices do not lose track of where one byte begins and another byte ends, if data is being sent through multiple lanes to a device, this data will usually be interleaved with sequential bytes being sent on different lanes, which further increases speed. PCIe speeds are typically in the 2.5 Gigabit/second ranges, with faster devices planned for the near future.
The transaction layer manages exactly what sort of traffic is moving over the serial lanes at any given Moment of time. It utilizes credit-based flow control. PCIe devices signals (get credit for) any extra receive buffers they may have. Whenever a sending device sends a transaction layer packet (here termed a PCIe-TLP to distinguish this from a different “thread level parallelism” TLP acronym) to a receiving device, it, the sending device deducts a credit from this account, thus ensuring that the buffer capability of the receiving device is not exceeded. When the receiving device has processed the data it sends a signal back to the sending device signaling that it has free buffers again. This way, a fair number of PCIe-TLP's can be reliably sent without cluttering bandwidth with a lot of return handshaking signals for each PCIe-TLP.
The data link layer handles the transaction layer packets. In order to detect any PCIe-TLP send or receive errors, the data link error bundles the PCIe-TLP with, a 32 bit CRC checksum. If, for some reason, a given PCIe-TLP fails checksum verification, this failure is communicated back to the originating device as a NAK using a separate type of data link layer packet (DLLP). The originating device can then retransmit the PCIe-TLP.
Further details of how the PCIe bus can be integrated with the advanced processor will be given later in this discussion.
In addition to the PCIe bus, the interface may contain many other types of devices. Also included, as part of the interface are two Reduced GMII (RGMII) (e.g., boxes 230a and 230b of
The architecture philosophy for the CPU is to optimize for thread level parallelism (TLP) rather than instruction level parallelism (ILP) including networking workloads benefit from TLP architectures, and keeping it small.
The architecture allows for many CPU instantiations on a single chip, which in turn supports scalability. In general, super-scalar designs have minimal performance wins on memory bound problems. An aggressive branch prediction is typically unnecessary for this type of processor app Heat ion and can even be wasteful.
The exemplary embodiment employs narrow pipelines because they typically have better frequency scalability. Consequently, memory latency is not as much of an issue as it would be in other types of processors, and in fact, any memory latencies can effectively be hidden by the multithreading as described below.
Embodiments of the invention can optimize the memory subsystem with non-blocking loads, memory reordering at the CPU interface, and special instruction for semaphores and memory barriers.
In one aspect of the invention, the processor can acquire and release semantics added to load/stores. In another aspect of embodiments of the invention, the processor can employ special atomic incrementing for timer support.
As described above, the multithreaded CPUs offer benefits over conventional techniques. An exemplary embodiment of the invention employs fine grained multithreading that can switch threads every clock and has 4 threads available for issue.
The multithreading aspect provides for the following advantages: usage of empty cycles caused by long latency operations; optimized for area versus performance trade-off; ideal for memory bound applications; enable optimal utilization of memory bandwidth; memory subsystem; cache coherency using MOSI (Modified, Own, Shared, invalid) protocol; full map cache directory including reduced snoop bandwidth and increased scalability over broadcast snoop approach; large on-chip shared dual banked 2 MB L2 cache; error checking and correcting (ECC) protected caches and memory; 2 64-bit 400/800 DDR2 channels (e.g., 12.8 Gbyte/s peak bandwidth) security pipeline; support of on-chip standard security functions (e.g., AES, DES/3DES, SHA-1, MD5, and RSA); allowance of the chaining of functions (e.g., encrypt->sign) to reduce memory accesses; 4 Gbs of bandwidth per security-pipeline, excluding RSA; on-chip switch interconnect; message passing mechanism for intra-chip communication; point-to-point connection between super-blocks to provide increased scalability over a shared bus approach; 16 byte full-duplex links for data messaging (e.g., 32 GB/s of bandwidth per link at 1 GHz); and credit-based flow control mechanism.
Some of the benefits of the multithreading technique used with the multiple processor cores include memory latency tolerance and fault tolerance.
The exemplary advanced processor 200 of
The exemplary processor includes the multiple CPU cores 210a-h capable of multithreaded operation. In the exemplary embodiment, there are eight 4-way multithreaded. MIPS64-compatible CPUs, which are often referred to as processor cores. Embodiments of the invention can include 32 hardware contexts and the CPU cores may operate at over approximately 1.5 GHz. One aspect of the invention is the redundancy and fault tolerant nature of multiple CPU cores. So, for example, if one of the cores failed, the other cores would continue operation and the system would experience only slightly degraded overall performance. In one embodiment, a ninth processor core may be added to the architecture to ensure with a high degree of certainty that eight cores are functional.
The multithreaded core approach can allow software to more effectively use parallelism that is inherent in many packet processing applications. Most conventional processors use a single-issue, single-threaded architecture, but this has performance limitations in typical networking applications. In aspects of the invention, the multiple threads can execute different software programs and routines, and even run different operating systems. This ability, similar to that described above with respect to the cores, to run different software programs and operating systems on different threads within a single unified platform can be particularly useful where legacy software is desired to be run on one or more of the threads under an older operating system, and newer software is desired to be run on one or more other threads under a different operating system or systems. Similarly, as the exemplary processor permits multiple separate functions to be combined within a unified platform, the ability to run multiple different software and operating systems on the threads means that the disparate software associated with the separate functions being combined can continue to be utilized. Discussed below are some techniques used by the invention to improve performance in single and multithreaded applications.
Referring now to
For many processors, performance is improved by executing more instructions per cycle, thus providing for instruction level parallelism (ILP). In this approach, more functional units are added to the architecture in order to execute multiple instruct ions per cycle. This approach is also known as a single-threaded, multiple-issue processor design. While offering some improvement over single-issue designs, performance typically continues to suffer due to the high-latency nature of packet processing applications in general. In particular, long-latency memory references usually result in similar inefficiency and increased overall capacity loss.
As an alternate approach, a multithreaded, single-issue architecture may be used. This approach takes advantage of, and more fully exploits, the packet level parallelism commonly found in net working applications. In short, memory latencies can be effectively hidden by an appropriately designed multithreaded processor. Accordingly, in such a threaded design, when one thread becomes inactive while waiting for memory data to return, the other threads can continue to process instructions. This can maximize processor use by minimizing wasted cycles experienced by other simple multi-issue processors.
Referring now to
Referring now to
Referring now to
Referring now to
Referring now to
Referring now to
According to embodiments of the invention, each MIPS architecture core may have a single physical pipeline, but may be configured to support multi-threading functions (i.e., four “virtual” cores). In a networking application, unlike a regular computational type of instruction scheme, threads are more likely to be waited on for memory accesses or other long latency operations. Thus, the scheduling approaches as discussed herein can be used to improve the overall efficiency of the system.
Referring now to
Viewing
Stages 5-10 of the example pipeline operation of
In general, the architecture is designed such that there are no stalls in the pipeline. This approach was taken for both ease of implementation as well as increased frequency of operation. However, there are some situations where a pipeline stall or stop is required. In such situations, Decoupling buffer 308G, which can be considered a functional part of IFU 304G, can allow for a restart or “replay” from a stop point instead of having to flush the entire pipeline and start the thread over to effect the stall. A signal can be provided by IFU 304G to Decoupling buffer 308G to indicate that a stall is needed, for example. In one embodiment, Decoupling buffer 308G can act as a queue for instructions whereby each instruction obtained by IFU 304G also goes to Decoupling buffer 308G. In such a queue, instructions may be scheduled out of order based on the particular thread scheduling, as discussed above. In the event of a signal to Decoupling buffer 3080 that a stall is requested, those instructions after the “stop” point can be re-threaded. On the other hand, if no stall is requested, instructions can simply be taken out of the decoupling buffer and the pipeline continued. Accordingly, without a stall, Decoupling buffer 308G can behave essentially like a first-in first-out (FIFO) buffer. However, if one of several threads requests a stall, the others can proceed through the buffer and they may not be held up.
As another aspect of embodiments of the invention, a translation-lookaside buffer (TLB) can be managed as part of a memory management unit (MMU), such as MMU 310G of
In order to maintain compliance with the MIPS architecture, the main TLB can support paired entries (e.g., a pair of consecutive virtual pages mapped to different physical pages), variable page sizes (e.g., 4K to 256 M), and software management via TLB read/write instructions. To support multiple threads, entries in the microTLB and in the main TLB may be tagged with the thread ID (TID) of the thread that installed them. Further, the main TLB can be operated in at least two modes. In a “partitioned” mode, each active thread may be allocated an exclusive subset or portion of the main TLB to install entries and, during translation, each thread only sees entries belonging to itself. In “global” mode, any thread may allocate entries in any portion of the main TLB and all entries may be visible to all threads. A “de-map” mechanism can be used during main TLB writes to ensure that overlapping translations are not introduced by different threads.
Entries in each microTLB can be allocated using a not-recently-used (NRU) algorithm, as one example. Regardless of the mode, threads may allocate entries in any part of the micro, LB. However, translation in the microTLB may be affected by mode. In global mode, all microTLB entries may be visible to all threads, but in partitioned mode, each thread may only see its own entries. Further, because the main TIE can support a maximum of one translation per cycle, an arbitration mechanism may be used to ensure that microTLB “miss” requests from all threads are serviced fairly.
In a standard MIPS architecture, unmapped regions of the address space follow the convention that the physical address equals the virtual address. However, according to embodiments of the invention, this restriction is lifted and unmapped regions can undergo virtual-to-physical mappings through the microTLB/mainTLB hierarchy while operating in a “virtual-MIPS” mode. This approach allows a user to isolate unmapped regions of different threads from one another. As a byproduct of this approach, however, the normal MIPS convention that mainTLB entries containing an unmapped address in their virtual page number (VPN2) field can be considered invalid is violated. In one embodiment of the invention, this capability can be restored to the user whereby each entry in the mainTLB can include a special “master valid.” bit that may only be visible to the user in the virtual MIPS-mode. For example, an invalid entry can be denoted by a master valid bit value of “0” and a valid entry can be denoted by a master valid bit value of “1.”
As another aspect of the invention, the system can support out-of-order load/store scheduling in an in-order pipeline. As an example implementation, there can be a user-programmable relaxed memory ordering model so as to maximize overall performance. In one embodiment, the ordering can be changed by user programming to go from a strongly ordered model to a weakly ordered model. The system can support four types: (i) Load-Load Re-ordering; (ii) Load-Store Re-ordering; (ii) Store-Store Re-ordering and (iv) Store-Load Re-ordering. Each type of ordering can be independently relaxed by way of a bit vector in a register. If each type is set to the relaxed state, a weakly ordered model can be achieved.
Referring now to
As shown in more detail in
Referring now to
Each interrupt and/or entry of IRT 308J can include associated attributes (e.g., Attribute 314J) for the interrupt, as shown. Attribute 314J can include CPU Mask 316-1J, Interrupt Vector 316-2J, as well as fields 316-3J and 316-4J, for examples. Interrupt Vector 316-2J can be a 6-bit field that designates a priority for the interrupt. In one embodiment, a lower number in Interrupt Vector 316-2J can indicate a higher priority for the associated interrupt via a mapping to EIPR 308I, as discussed above with reference to
In addition to the PIC, each of 32 threads, for example, may contain a 64-bit interrupt vector. The PIC may receive interrupts or requests from agents and then deliver them to the appropriate thread. As one example implementation, this control may be software programmable. Accordingly, software control may elect to redirect all external type interrupts to one or more threads by programming the appropriate PIC control registers. Similarly, the PIC may receive an interrupt event or indication from the PCI-X interface (e.g., PCI-X 234 of
In the case where multiple recipients are programmed for a given event, or interrupt, the PIC scheduler can be programmed to use a global “round-robin” scheme or a per-interrupt-based local round-robin scheme for event delivery. For example, if threads 5, 14, and 27 are programmed to receive external interrupts, the PIC scheduler may deliver the first external interrupt to thread 5, the next one to thread 14, the next one to thread 27, then return to thread 5 for the next interrupt, and so on.
In addition, the PIC also may allow any thread to interrupt any other thread (i.e., an inter-thread interrupt). This can be supported by performing a store (i.e., a write operation) to the PIC address space. The value that can be used for such a write operation can specify the interrupt vector and the target thread to be used by the PIC for the inter-thread interrupt. Software control can then use standard conventions to identify the inter-thread interrupts. As one example, a vector range may be reserved for this purpose.
As discussed above with reference to
In one embodiment of the invention, a 3-cycle cache can be used in the core implementation. Such a 3-cycle cache can be an “off-the-shelf” cell library cache, as opposed to a specially-designed cache, in order to reduce system costs. As a result, there may be a gap of three cycles between the load and the use of a piece of data and/or an instruction. The decoupling buffer can effectively operate in and take advantage of this 3-cycle delay. For example, if there was only a single thread, a 3-cycle latency would be incurred. However, where four threads are accommodated, intervening slots can be taken up by the other threads. Further, branch prediction can also be supported. For branches correctly predicted, but not taken, there is no penalty. For branches correctly predicted and taken, there is a one-cycle “bubble” or penalty. For a missed prediction, there is a 5-cycle bubble, but such a penalty can be vastly reduced where four threads are operating because the bubbles can simply be taken up by the other threads. For example, instead of a 5-cycle bubble, each of the four threads can take up one so that only a single bubble penalty effectively remains.
As discussed above with reference to
Referring now to
In the example operation of
For multiple thread operation. Stack 312K can be partitioned so that entries are dynamically configured across a number of threads. The partitions can change to accommodate the number of active threads. Accordingly, if only one thread is in use, the entire set of entries allocated for Stack 312K can be used for that thread. However, if multiple threads are active, the entries of Stack 312K can be dynamically configured to accommodate the threads so as to utilize the available space of Stack 312K efficiently.
In a conventional multiprocessor environment, interrupts are typically given to different CPUs for processing on a round-robin basis or by designation of a particular CPU for the handling of interrupts. However, in accordance with embodiments of the invention, PIC 226 of
Further, for both CPUs/cores as well as threads, a round-robin scheme (e.g., by way of a pointer) can be employed among those cores and/or threads that are not masked for a particular interrupt. In this fashion, maximum programmable flexibility is allowed for interrupt load balancing. Accordingly, operation 300j of
As another aspect of embodiments of the invention, thread-to-thread interrupting is allowed whereby one thread can interrupt, another thread. Such thread-to-thread interrupting may be used for synchronization of different threads, as is common for telecommunications applications. Also, such thread-to-thread interrupting may not go through any scheduling according to embodiments of the invention.
Returning now to
As previously discussed, embodiments of the invention may include the maintenance of cache coherency using MOSI (Modified, Own, Shared, Invalid) protocol. The addition of the “Own” state enhances the “MSI” protocol by allowing the sharing of dirty cache lines across process cores. In particular, an example embodiment of the invention may present a fully coherent view of the memory to software that may be running on up to 32 hardware contexts of 8 processor cores as well as the I/O devices. The MOSI protocol may be used throughout the L1 and L2 cache (e.g., 212a-h and 208, respectively, of
According to one aspect of embodiments of the invention, an L2 cache (e.g., cache 208 of
As to ECC protection for an L2 cache implementation, both cache data and cache tag arrays can be protected by SECDED (Single Error Correction Double Error Detection) error protecting codes. Accordingly, all single bit errors are corrected without software intervention. Also, when uncorrectable errors are detected, they can be passed to the software as code-error exceptions whenever the cache line is modified. In one embodiment, as will be discussed in more detail below, each L2 cache may act like any other “agent” on a ring of components.
According to another aspect of embodiments of the invention, “bridges” on a data Movement ring may be used for optimal redirection of memory and I/O traffic. Super Memory I/O Bridge 206 and Memory Bridge 218 of
Referring now to
As shown in
Referring now to
Incoming data or “Ring In” can be received in flip-flop 404B. An output of flip-flop 40413 can connect to flip-flops 40613 and 408B as sell as multiplexer 41613. Outputs of flip-flops 40613 and 40813 can be used for local data use. Flip-flop 41013 can receive an input from the associated L2 cache while flip-flop 412B can receive an input from the associated CPU. Outputs from flip-flops 410B and 412B can connect to multiplexer 414B. An output of multiplexer 414B can connect to multiplexer 416B and an output of multiplexer 416B can connect to outgoing data or “Ring Out.” Also, ring component 402b-0 can receive a valid bit signal.
Generally, higher priority data received, on Ring In will be selected by multiplexer 416B if the data is valid (e.g., Valid Bit=“1”). If not, the data can be selected from either the L2 or the CPU via multiplexer 414B. Further, in this example, if data received on Ring In is intended for the local node, flip-flops 406B and/or 408B can pass the data onto the local core instead of all lowing the data to pass all the way around the ring before receiving it again.
Referring now to
In an alternative embodiment, the memory bridge would not have to wait for an indication that the data has not been found in any of the L2 caches in order to initiate the memory request. Rather, the memory request (e.g., to DRAM), may be speculatively issued. In this approach, if the data is found prior to the response from the DRAM the later response can be discarded. The speculative DRAM accesses can help to mitigate the effects of the relatively long memory latencies.
Also in
Referring now to
Referring now to
The credit-based flow control can provide a mechanism for man aging message sending, for example. In one embodiment, the total number of credits assigned to all transmitters for a target/receiver cannot exceed the sum of the number of entries in its receive queue (e.g., RCVQ 506A of
In this example, when Core-1 sends a message of size 2 (e.g., 2 64-bit data elements) to Core-0, the Core-1 credit in Core-0 can be decremented by 2 (e.g., from 32 to 30). When Core-0 receives a message, the message can go into the RCVQ of Core-0. Once the message is removed from the RCVQ of Core-0, that message storage space may essentially be freed-up or made available. Core-0 can then send a signal to the sender (e.g., a free credit signal to Core-1) to indicate the amount of space (e.g., 2) additionally available. If Core-1 continues to send messages to Core-0 without corresponding free credit signals from Core-0, eventually the number of credits for Core-1 can go to zero and Core-1 may not be able to send any more messages to Core-0. Only when Core-0 responds with free credit signals could Core-1 send additional messages to Core-0, for example.
Referring now to
As an aspect of embodiments of the invention, all agents (e.g., cores/threads or networking interfaces, such as shown in
In another aspect of embodiments of the invention, all threads of the core e.g. Core-0502C-0 through Core-7502C-7 or
Referring now to
Referring now to
In one aspect of embodiments of the invention, the FMN can interface to each CPU/core, as shown in
As another aspect of embodiments of the invention, fast messaging (FMN) ring components can be organized into “buckets.” For, example, RCVQ 506A and XMTQ 508A of
In one aspect of embodiments of the invention, a Packet Distribution Engine (PDE) can include each of the XGMII/SPI-4.2 interfaces and four RGMII interfaces and PCIe interfaces to enable efficient and load-balanced distribution of incoming packets to the processing threads. Hardware accelerated packet distribution is important for high throughput networking applications. Without the PDF, packet distribution may be handled by software, for example. However, for 64 B packets, only about 20 ns is available for execution of this function on an XGMII type interface. Further, queue pointer management would have to be handled due to the single-producer multiple-consumer situation. Such a software-only solution is simply not able to keep up with the required packet delivery rate, without impacting the performance of the overall system.
According to an embodiment of the invention, the PDE can utilize the Fast Messaging Network (FMN) to quickly distribute packets to the threads designated by software as processing, threads. In one embodiment, the PDE can implement a weighted round-robin scheme for distributing packets among the intended recipients. In one implementation, a packet is not actually moved, but rather gets written to memory as the networking interface receives it. The PDE can insert a “Packet Descriptor” in the message and then send it to one of the recipients, as designated by software. This can also mean that not all threads must participate in receiving packets from any given interface.
Referring now to
To speed development of such devices, it will usually be advantageous to rely upon pre-designed PCIe components as much as possible. Suitable PCIe digital cores, physical layer (PHY), and verification components may be obtained from different vendors. Typically these components will be purchased as intellectual property and integrated circuit design packages, and these design packages used, in conjunction with customized DMA design software, to design integrated circuit chips capable of interfacing to fast message buses of advanced processors, and in turn interfacing with PCIe bus hardware.
One important design consideration is to streamline (simplify) the PCIe interface and commands as much as possible. This streamlining process keeps the interface both relatively simple and relatively fast, and allows single chip multiple core processors, such as the present invention, to control PCIe devices with relatively small amounts of software and hardware overhead.
As will be discussed, the present design utilizes the Fast messaging network (FMN)/interface switch interconnect in combination with a customized DMA engine (541) embedded as part attic PCIe interface unit (534). The DMA engine essentially serves as a translator between the very different memory storage protocols used by the processor cores and the various PCIe devices. Offloading the translation process to a customized DMA engine greatly reduces the computing demands on the core processors, freeing them up for other tasks.
To briefly review, DMA (Direct hi empty Access) circuits allow hardware to access memory independently of the processor core CPUs. The DMA acts to copy memory chunks between devices. Although the CPU initiates this process with an appropriate DMA command, the CPU can then do other things while the DMA executes the memory copy command.
In this embodiment, a customized. DMA engine was constructed that accepted the short, the highly optimized, processor FMN and ISI messages as input. The DMA then both translated the FMN/ISI messages into appropriate PCIe-TLP packets (which were then handled by the other PCIe circuitry in the interface), and then automatically handled the original memory copy command requested by the core processors.
Thus the DMA translated the memory copy request into PCIe-TLP packets, and sent it on to the other PCIe circuitry that sends the appropriate commands to the appropriate PCIe device on the PCIe bus via PCIe-TLP packets. After these packets had been sent by the PCIe circuitry, the designated PCIe device completes the task. After the designated PCIe bus device has completed its assigned task, the PCIe bus device would then return appropriate PCIe-TLP packets to the PCIe interface. The PCIe interface DMA accepts these PCIe-TLP packets, and translates them into the appropriate memory copy commands, and otherwise manages the task originally assigned by the processor core without need for further attention from the processor cores. As a result, the processor and the processor cores can now communicate with a broad variety of different PCIe devices with minimal processor core computational overhead.
The processor cores communicate with the PCIe interface via short (1-2 64 bit words) messages sent over the FMN/ISI. The PCIe interface is designed to accept incoming messages that are assigned to four different buckets. Each bucket can have a queue of up to 64 entries per bucket, for a maximum queue dept of 256 messages.
The PCIe interface is also designed to output 4 different classes of messages. These output messages can each be stored in a queue of up to 4 messages per class, for a maximum output queue depth of 16 output messages.
The commands sent from the FMN/ISI to the PCIe interface are as follows:
Most of the communication from the processor cores to the PCIe interface is in the form of simple, terse, 2 64-bit word memory read/write commands. The processor cores will normally assume that these commands have been successfully completed. As a result, the PCIe interface communicates back only terse OK, Not OK, status messages in the form of a very terse single 64-bit word. The various message out classes are shown below:
The fields of the FMN/ISI message to the PCIe interface DMA engine are as follows
Here, the COHERENT, RDX, RETEN, TD, EP ATTR fields can be considered to be star topology serial bus control fields, or more specifically PCIe bus control fields.
The outgoing return message is essentially a glorified ACK or OK or problem message. It is composed of a single 64 bit word. Its fields are shown below:
To simplify this discussion, the prowess of these return messages will not be discussed in detail, however some of this return path is shown indicated, in
Returning to
The messages are then processed through a round robin scheduler and a message bus arbiter to the DMA engine (541). Memory load (write) instructions are processed by the DMA load posted requestor portion of the DMA engine (543), translated into appropriate PCIe-TCP data packets, and are then sent to the I/O bus slave circuit (545). This is then sent to the I/O bus arbiter circuit (546), which handles various PCI I/O bus requests using a round robin scheduler. The I/O bus arbiter then sends the packets to other PCIe circuitry, such as the I/O bus requestor (547) and the PCIe link completors (548). These work with the PCIe control arbiter (549) to manage the actual PCIe physical layer circuitry (550) that does the actual hardware PCIe packet sending and receiving. The PCIe-TCP packets then are put on the PCIe bus (551) and sent to the various PCIe devices (not shown). The write tag tracker tracks the status of reads out and content in.
In this embodiment, all write commands do not expect responses and are complete once the commands are accepted, however all read and read exclusive commands expect return responses.
In this embodiment, in order to simplify the design, it is assumed that all I/O bus master read request completions can be returned out of order. In this embodiment, it is also assumed that all requests are for contiguous bytes, and that for the I/O, configuration, and register address spaces, the most number of byes requested in a command is 4 and the bytes do not cross the Double Word address boundary.
As previously discussed, to simplify the system and make processing time faster, in this embodiment, the length of the write data plus the destination address double word offset cannot exceed the maximum pay load length of the PCIe-TLP. Further, the write request cannot cross 4K-byte address boundaries. The PCIe-TLPs are further divided into 32-byte source addresses
Return PCIe-TLP data packets: Memory read (store) requests (again the two 64 bit words shown in
The DMA store requestor return packet handler (567) will make the appropriate writes via the I/O (559) to cache or dram (558), consult the original DMA instructions, as well as the contents of the PCIe-TLP packets, and either generate additional PCIe-TLP packets to forward the data on to the appropriate PCIe device. It will then notify the processor cores (CPU) when read or write to the DRAM is complete.
The PCIe physical layer receiver receives (566) receives PCIe-TLP packets (567) from various PCIe devices. The DMA store requestor engine determines if the appropriate way to respond to the original DMA command is or is not to send out additional PCIe-TLP packets to other PCIe devices. If so, it sends appropriate PCIe-TMP commands out again. If the appropriate response to the received PCIe-TLP packets and the original DMA command is to write the data to L2 cache or place it back on the FMN/ISI bus, the DMA will again make the appropriate decision and route it to the FMN/ISI as needed.
Referring now to
Referring now to
Because most net working applications are not very tolerant of the random arrival order of packets, it is desirable to deliver packets in order. In addition, it can be difficult to combine features of parallel processing and packet ordering in a system. One approach is to leave the ordering task to software, but it then becomes difficult to maintain line rate. Another option is to send all packets in a single flow to the same processing thread so that the ordering is essentially automatic. However, this approach would require flow identification (i.e., classification) prior to packet distribution and this reduces system performance. Another drawback is the throughput of the largest flow is determined by the performance of the single thread. This prevents single large flows from sustaining their throughput as they traverse the system.
According to an embodiment of the invention, an advanced hardware-accelerated structure called a Packet Ordering Device (POD) can be used. An objective of the POD is to provide an unrestricted use of parallel processing threads by re-ordering the packets before they are sent to the networking output interface. Referring now to
Packet Ordering Device (POD) 604C can be considered a packet sorter in receiving the packets from the different threads and the sending to Networking Output. All packets received by a given networking interface can be assigned a sequence number. This sequence number can then be forwarded to the working thread along with the rest of the packet information, by the PDE. Once a thread has completed processing the packet, it can forward the packet descriptor along with the original sequence number to the POD. The POD can release these packets to the outbound interface in an order strictly determined by the original sequence numbers assigned by the receiving interface, for example.
In most applications, the POD will receive packets in a random order because the packets are typically processed by threads in a random order. The POD can establish a queue based on the sequence number assigned by the receiving interface and continue sorting packets as received. The POD can issue packets to a given outbound interface in the order assigned by the receiving interface. Referring now to
It is possible that the oldest packet may never arrive in the POD, thus creating a transient head-of-line blocking situation. If not handled properly, this error condition would cause the system to deadlock. However, according to an aspect of the embodiment, the POD is equipped with a time-out mechanism designed to drop a non-arriving packet at the head of the list once a time-out counter has expired. It is also possible that packets are input to the POD at a rate which fills the queue capacity (e.g. 512 positions) before the time-out counter has expired. According to an aspect of the embodiment, when the POD reaches queue capacity, the packet at the head of the list can be dropped and a new packet can be accepted. This action may also remove any head-of-line blocking situation as well. Also, software may be aware that a certain sequence number will not be entered into the POD due to a bad packet, a control packet, or some other suitable reason. In such a case, software control may insert a “dummy” descriptor in the POD to eliminate the transient head-of-line blocking condition before allowing the POD to automatically react.
According to embodiments of the invention, five programmable PODs may be available (e.g., on chip) and can be viewed as generic “sorting” structures. In one example configuration, software control (i.e., via a user) can assign four of the PODs to the four networking interfaces while retaining one POD for generic sorting purposes. Further, the PODs can simply be bypassed if so desired for applications where software-only control suffices.
In one aspect of embodiments of the invention, the advanced telecommunications processor can further include memory bridge 218 coupled to the data switch interconnect and at least one communication port (e.g., box 220), and configured to communicate with the data switch interconnect and the communication port.
In one aspect of the invention, the advanced telecommunications processor can further include super memory bridge 206 coupled to the data switch interconnect (DSI), the interface switch interconnect and at least one communication port (e.g., box 202, box 204), and configured to communicate with the data switch interconnect, the interface switch interconnect and the communication port.
In another aspect of embodiments of the invention, memory ordering can be implemented on a ring-based data movement network, as discussed above with reference to
Advantages of the invention include the ability to provide hilt bandwidth communications between computer systems and memory in an efficient and cost-effective manner. In particular, this embodiment of the invention focuses on a novel PCIe interface that enhances the capability of the advanced processor by enabling the processor to work with a variety of different PCIe devices.
Having disclosed exemplary embodiments and the best mode, modifications and variations may be made to the disclosed embodiments while remaining within the subject and spirit of the invention as defined by the following Claims.
This application is a continuation of Ser. No. 11/831,887 filed Jul. 31, 2007, which is a continuation in part of Ser. No, 10/930,937 filed Aug. 31, 2004, which is a continuation in part of Ser. No. 10/898,008 filed Jul. 23, 2004, which is a continuation in part of Ser. No. 10/682,579 filed Oct. 8, 2003, which claimed priority to Prov. No. 60/490,236 filed Jul. 25, 2003 and Prov. No. 60/416,838 filed Oct. 8, 2002, all incorporated herein by reference and all priorities claimed.
| Number | Date | Country | |
|---|---|---|---|
| 60490236 | Jul 2003 | US | |
| 60416838 | Oct 2002 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 12019576 | Jan 2008 | US |
| Child | 12815092 | US | |
| Parent | 11831887 | Jul 2007 | US |
| Child | 12019576 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10930937 | Aug 2004 | US |
| Child | 11831887 | US | |
| Parent | 10898008 | Jul 2004 | US |
| Child | 10930937 | US | |
| Parent | 10682579 | Oct 2003 | US |
| Child | 10898008 | US |
