Method, apparatus, and program for eliminating thread skew in multithreaded performance benchmarks

• US 7,257,516 B2
• Filed: 09/20/2001
• Issued: 08/14/2007
• Est. Priority Date: 09/20/2001
• Status: Expired due to Fees

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: International Business Machines Corporation
• Inventors: Arora, Rajiv, Gordon, Jesse Mannes
• Primary Examiner(s): Tsai; Carol S. W.
• Application Number:
  US09/956,768
• Publication Number:
  US 20030055764A1
• Time in Patent Office: 2,154 Days
• Field of Search: 713/1, 713/100, 713/300, 709/224, 709/220, 709/221, 718/100, 718/101, 718/102, 703/17, 711/137, 714/38, 714/47, 717/124, 702/186, 370/232, 370/229
• US Class Current: 702/186
• CPC Class Codes: G06F 11/3428 Benchmarking
• CPC Class Codes: G06Q 20/40 Authorisation, e.g. identification of payer or payee, verification of customer or shop credentials; Review and approval of payers, e.g. check credit lines or negative lists
• CPC Class Codes: G06Q 40/06 Asset management; Financial planning or analysis

Abstract

A data processing system performs a multithreaded performance benchmark with a rampup interval and a rampdown interval. The master thread signals a start test event and begins the rampup interval. After the rampup interval, the master thread signals a start measurement event. In response to the start measurement event, the worker threads record the units of work they complete. After the measurement interval, the master signals the workers to stop measuring, but to continue running. This begins the rampdown interval. After the rampdown interval, the master thread signals a stop test event. The rampup and rampdown intervals are long enough to ensure that measurements are not recorded during skew intervals. Thus, thread skew does not impact the results.

Claims

1. A data processing system implemented method for performing a benchmark, comprising steps implemented by the data processing system of:

signaling a start test event and waiting during a rampup interval;

signaling a start measurement event at the end of the rampup interval;

responsive to the signaling the start measurement event, collecting measurement results for units of work performed by a plurality of worker threads, the measurement results having mitigated thread skew due to the measurement results being collected responsive to the signaling the start measurement event; and

signaling a stop test event.

2. The method of claim 1, wherein events are signaled by setting a shared memory flag.

3. The method of claim 1, wherein events are signaled using operating system signals.

4. The method of claim 1, further comprising starting plurality of worker threads during the rampup interval.

5. The method of claim 4, wherein the step of collecting measurement results comprises:

signaling a stop measurement event; and

receiving the measurement results from the worker threads.

6. The method of claim 5, further comprising:

waiting during a rampdown interval after the stop measurement event; and

signaling the stop test event at the end of the rampdown interval.

7. The method of claim 6, wherein the length of the rampdown interval is determined based on the number of worker threads in the plurality of worker threads.

8. A computer readable medium encoded with a computer program product data structure adapted to perform the steps of claim 1, when said computer program product is run on a data processing system.

9. A data processing system implemented method for performing a benchmark, comprising steps implemented by the data processing system of:

signaling a start test event;

starting a plurality of worker threads during a rampup interval;

signaling a start measurement event at the end of the rampup interval;

collecting measurement results; and

signaling a stop test event, wherein the length of the rampup interval is determined based on the number of worker threads in the plurality of worker threads.

10. A data processing system implemented method for performing a benchmark, comprising steps implemented by the data processing system of:

signaling a start test event;

waiting during a rampup interval;

signaling a start measurement event at the end of the rampup interval;

collecting measurement results;

signaling a stop test event;

repeatedly increasing the rampup interval and performing the benchmark until the increase of the rampup interval has no measurable impact on the measurement results.

11. A data processing system implemented method for performing a benchmark, comprising steps implemented by the data processing system of:

signaling a start test event;

waiting during a rampup interval;

signaling a start measurement event at the end of the rampup interval;

collecting measurement results, wherein the step of collecting measurement results comprises repeatedly running an iteration of a test until a plurality of metrics have converged; and

signaling a stop test event.

12. The method of claim 11, further comprising, for each iteration of the test:

computing a plurality of metrics;

computing confidence intervals for each of the plurality of metrics; and

determining whether each confidence interval is less than a respective threshold, wherein the plurality of metrics have converged if all confidence intervals are less than their respective thresholds.

13. A data processing system implemented method for performing a benchmark, comprising:

starting a plurality of worker threads;

waiting during a rampup interval;

signaling a start measurement event at the end of the rampup interval;

responsive to the signaling the start measurement event, receiving measurement results from the worker threads, the measurement results having mitigated thread skew due to the measurement results being received responsive to the signaling the start measurement event;

signaling a stop measurement event;

waiting during a rampdown interval after the stop measurement event; and

signaling a stop test event at the end of the rampdown interval.

14. A computer readable medium encoded with a computer program product data structure adapted to perform the steps of claim 13, when said computer program product is run on a data processing system.

15. A data processing machine for performing a benchmark, comprising:

first signal means for signaling a start test event and waiting during a rampup interval by the data processing machine;

second signal means for signaling a start measurement event at the end of the rampup interval by the data processing machine;

collection means, responsive to the second signaling means, for collecting measurement results comprising units of work performed by a plurality of worker threads by the data processing machine; and

third signal means for signaling a stop test event by the data processing machine.

16. The data processing machine of claim 15, wherein events are signaled by setting a shared memory flag within the data processing machine.

17. The data processing machine of claim 15, wherein events are signaled using operating system signals within the data processing machine.

18. The data processing machine of claim 15, further comprising means for starting the plurality of worker threads during the rampup interval by the data processing machine.

19. The data processing machine of claim 15, wherein the collection means comprises:

means for signaling a stop measurement event by the data processing machine; and

means for receiving the measurement results from the worker threads by the data processing machine.

20. The data processing machine of claim 19, further comprising:

means for waiting during a rampdown interval after the stop measurement event by the data processing machine; and

means for signaling the stop test event at the end of the rampdown interval by the data processing machine.

21. The data processing machine of claim 15, further comprising:

means for repeatedly increasing the rampup interval and performing the benchmark by the data processing machine until the increase of the rampup interval has no measurable impact on the measurement results.

22. The data processing machine of claim 15, wherein the collection means comprises:

iteration means for repeatedly running an iteration of a test by the data processing machine until a plurality of metrics have converged.

23. A data processing machine for performing a benchmark, comprising:

means for starting a plurality of worker threads by the data processing machine;

means for waiting during a rampup interval by the data processing machine;

means for signaling a start measurement event at the end of the rampup interval by the data processing machine;

means, responsive to the signaling the start measurement event, for receiving measurement results from the worker threads by the data processing machine;

means for signaling a stop measurement event by the data processing machine;

means for waiting during a rampdown interval after the stop measurement event by the data processing machine; and

means for signaling a stop test event at the end of the rampdown interval by the data processing machine.