1. Field of the Invention
The present invention is related to performance measurements in processing systems and processors. In particular, the present invention relates to a processor core having a saturating event counter for making processor performance measurements.
2. Description of Related Art
Performance measurements are used in both system component and software evaluation, as well as in run-time applications such as workload balancing, resource usage accounting and other functions in computer systems in which a measurement of the efficiency or throughput of a workload executing within the computer system is needed. Performance monitoring can be performed in both hardware and software in order to measure and monitor performance of the system. Performance-related events, which may be events indicative of low performance such as cache misses, thread stalls, and the like, or may be events indicative of high performance such as instruction completions or instruction dispatches, can be detected to provide a indication of performance of a system, software or a particular component of a system such as a processor core.
Performance monitors can be implemented using counters that count occurrences of events having a frequency indicative of performance of the computer system. A counter-based event monitoring approach typically requires frequency monitoring of the performance counter count values, so that overflow of the counters and consequent wrap-around due to an occurrence of a large number of performance events is not missed. In such monitors, counter overflow may be allowed to occur and is taken into account, or overflow can be prevented by resetting the counter at a periodic rate that ensures that overflow will not occur, e.g., when the counter is read. However, either of the above-described approaches can lead to an erroneous condition, in which a low count value results when, in fact, a large number of performance events have occurred in the preceding measurement interval. Additionally, typical performance counter implementations do not readily provide information about temporal distributions of performance-related events, as it is difficult to identify when a previous performance-related event occurred in relation to the time of occurrence of another performance-related event.
Therefore, it would be desirable to provide performance monitoring in a computer system that does not require management of counter overflows or that requires periodically resetting the counters.
The above objectives, as well as others, are accomplished in a performance monitoring method and system that use saturating counters to provide an indication of the relative frequency of occurrence of performance related-events. The method may be embodied in a processor core that includes one or more saturating counters that measure the relative rate of internal performance-related event occurrences.
Events indicative of performance within a processor are detected and counted during predetermined intervals. When an event is detected, a saturating counter is incremented. If an event is not detected within the predetermined interval, then the saturating counter is decremented. The saturating counter thereby provides count value that indicates a relative frequency of performance-related events. The saturating counter or counters are then read to obtain a count value from which a relative performance level of the system can be determined.
A saturating counter can be maintained for each of a number of performance event types being detected for performance monitoring purposes. The predetermined period can be determined by a real time clock tick, a processor clock cycle, or by instruction cycles. The saturating counter can be of any bit width and software can read the counter at any time without overflow error. Instruction address and/or data address capture can be performed in response to counter saturation or a particular count level to identify program code sections, data values or time periods associated with a high frequency of performance-related events.
The foregoing and other objectives, features, and advantages of the invention will be apparent from the following, more particular, description of the preferred embodiment of the invention, as illustrated in the accompanying drawings.
The novel features believed characteristic of the invention are set forth in the appended claims. The invention itself, however, as well as a preferred mode of use, further objectives, and advantages thereof, will best be understood by reference to the following detailed description of the invention when read in conjunction with the accompanying Figures, wherein like reference numerals indicate like components, and:
The present invention relates to processors and processing systems in which performance is measured by counting performance-related events, i.e., events indicative of system performance, using a saturating counter. The saturating counter is incremented in response to detecting a performance related event, and is periodically decremented when no event is detected in a preceding period. Performance measurement in accordance with an embodiment of the invention can be used to determine a relative rate of processor performance on a temporal basis. For example, due to the periodic decrementing of the saturating counter, the performance count value will indicate the relative amount of performance-related events that have occurred within a periods of time related to the period of the decrementing. Also, the saturating counter performance counter of the present invention prevents the loss or misrepresentation of performance data that can otherwise result from a conventional performance counter that overflows.
Referring now to
Referring now to
In illustrated core 20, signals indicative of occurrences of performance-related events are provided to a performance monitor 42. Exemplary events in illustrated core 20 include misses in L1 I-cache 21A, L1 D-cache 21B and L2 cache 23, instruction dispatches by GDU 25, exceptions from FPU 31 and instruction completions from GCU 34. The illustrated event types and associated event signals are only exemplary and other types of events that are indicative of performance may be detected within core 20 and used to provide performance monitor 42 with performance monitoring input. Performance monitor 42 accumulates events according to their occurrence, which may be asynchronous or according to a periodic detection cycle. Performance monitor 42 detects performance-related events and increments a saturating counter according to the number of events detected. Performance monitor 42 also decrements the counter in response to the absence of a performance-related event occurring during a predetermined period. The saturating counter thus provides a temporal indication of performance of the processor in that higher count values will occur during periods of higher event rates. The events may be indicative of desirable performance, or undesirable performance and when multiple event types are combined in a performance monitor result, relative importance is weighted and the negative/positive aspects handled by appropriate sign of the individual contributions. For example, cache misses are generally negative performance events, while instruction dispatches and completions are generally positive performance events.
Referring now to
Saturating counter 45 may also be reset in response to a reset signal provided from an optional reset timer 47 that has a period set by a period reset value. The output of saturating counter 45 is provided to output register 46, which receives the count value of saturating counter 45 as an output count value, providing read access to the count value of saturating counter either by a memory or I/O mapped register value, or via another access mechanism such as a service control port or scan latch access. The count value output of output register 46 can be used in computing indications of processor workload or performance that can be displayed via a user interface, and as mentioned above, may comprise multiple values associated with multiple event types, provided by multiple corresponding saturating counters 45 that are effectively duplicates of the circuits described above. Control logic 44A may also control reset timer 47 according to a selectable operating mode, such that in a first operating mode, saturating counter is decremented only according to the periodic interval determined by periodic counter 44B, but in another operating mode, reset timer 47 periodically resets saturating counter at a period that is independent of the period of periodic counter 44B, although independence is not a requirement of such an embodiment of the invention, and reset tinier 47, may, for example, have a period that is a multiple of the period of periodic counter 44B.
Referring now to
Referring now to
While the invention has been particularly shown and described with reference to the preferred embodiments thereof, it will be understood by those skilled in the art that the foregoing and other changes in form, and details may be made therein without departing from the spirit and scope of the invention.
Number | Name | Date | Kind |
---|---|---|---|
5557548 | Gover et al. | Sep 1996 | A |
5657253 | Dreyer et al. | Aug 1997 | A |
5796637 | Glew et al. | Aug 1998 | A |
6112318 | Jouppi et al. | Aug 2000 | A |
6154857 | Mann | Nov 2000 | A |
6233531 | Klassen et al. | May 2001 | B1 |
6275782 | Mann | Aug 2001 | B1 |
6351724 | Klassen et al. | Feb 2002 | B1 |
6360337 | Zak et al. | Mar 2002 | B1 |
6546359 | Week | Apr 2003 | B1 |
6678777 | Rao et al. | Jan 2004 | B2 |
6792392 | Knight | Sep 2004 | B1 |
6804770 | Logan et al. | Oct 2004 | B2 |
6898261 | Hunter et al. | May 2005 | B1 |
7069426 | Hummel | Jun 2006 | B1 |
7284158 | Hunter et al. | Oct 2007 | B2 |
7340378 | Floyd | Mar 2008 | B1 |
7475295 | Purrington et al. | Jan 2009 | B2 |
7519510 | Koehler et al. | Apr 2009 | B2 |
7533003 | Floyd et al. | May 2009 | B2 |
7681054 | Ghiasi et al. | Mar 2010 | B2 |
20020073255 | Davidson et al. | Jun 2002 | A1 |
20020138713 | Logan et al. | Sep 2002 | A1 |
20040123185 | Pierce et al. | Jun 2004 | A1 |
20040216113 | Armstrong et al. | Oct 2004 | A1 |
20040236993 | Adkisson et al. | Nov 2004 | A1 |
20040237003 | Adkisson et al. | Nov 2004 | A1 |
20050081101 | Love et al. | Apr 2005 | A1 |
20050155021 | DeWitt et al. | Jul 2005 | A1 |
20050177344 | Khaleel | Aug 2005 | A1 |
20050183065 | Wolczko et al. | Aug 2005 | A1 |
20050246590 | Lancaster | Nov 2005 | A1 |
20050283677 | Adkisson et al. | Dec 2005 | A1 |
20080040634 | Matsuzaki et al. | Feb 2008 | A1 |
20080163254 | Cota-Robles | Jul 2008 | A1 |
Entry |
---|
May, John M. , “MPX: Software for Multiplexing Hardware Performance Counters”, IPDPS Proceedings 2001. |
Ghiasi, et al., “Scheduling for Heterogeneous Processors in Server Systems”, Conf. on Computing Frontiers, pp. 199-210, May 2005. |
Kotla, et al., “Scheduling Processor Voltage and Frequency in Server and Cluster Systems”, IPDPS 19th Symposium, Apr. 2005. |
Kotla, et al., “Characterizing the Impact of Different Memory Intensity Levels”, 7th Annual Workshop on Workload Characterization, Oct. 25, 2004. |
Number | Date | Country | |
---|---|---|---|
20120046912 A1 | Feb 2012 | US |