MONITORING SYSTEM NOISES IN PARALLEL COMPUTER SYSTEMS

• US 20150026687A1
• Filed: 07/18/2013
• Published: 01/22/2015
• Est. Priority Date: 07/18/2013
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: SEELAM, Seetharami R., FONG, Liana L., IYENGAR, Arun, LEWARS, John, YIM, Keun Soo
• Granted Patent: US 9,361,202 B2
• Time in Patent Office: Days
• US Class Current: 718/102
• CPC Class Codes: G06F 11/3024 where the computing system component is a central processing unit [CPU]
• CPC Class Codes: G06F 11/3072 where the reporting involves data filtering, e.g. pattern matching, time or event triggered, adaptive or policy-based reporting
• CPC Class Codes: G06F 11/3404 for parallel or distributed programming
• CPC Class Codes: G06F 11/3409 for performance assessment
• CPC Class Codes: G06F 11/3419 by assessing time
• CPC Class Codes: G06F 11/3476 Data logging G06F11/14, G06F11/2205 take precedence
• CPC Class Codes: G06F 2201/815 Virtual
• CPC Class Codes: G06F 2201/88 Monitoring involving counting
• CPC Class Codes: G06F 9/4881 Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
• CPC Class Codes: G06F 9/522 Barrier synchronisation

Abstract

Various embodiments monitor system noise in a parallel computing system. In one embodiment, at least one set of system noise data is stored in a shared buffer during a first computation interval. The set of system noise data is detected during the first computation interval and is associated with at least one parallel thread in a plurality of parallel threads. Each thread in the plurality of parallel threads is a thread of a program. The set of system noise data is filtered during a second computation interval based on at least one filtering condition creating a filtered set of system noise data. The filtered set of system noise data is then stored.

Claims

1. A method for monitoring system noise in a parallel computing system, the method comprising:

storing, during a first computation interval, at least one set of system noise data in a buffer, wherein the set of system noise data is detected during the first computation interval and is associated with at least one thread in a plurality of threads, and wherein each thread in the plurality of threads is a thread of a program;

filtering, during a second computation interval, the set of system noise data based on at least one filtering condition, the filtering creating a filtered set of system noise data; and

storing the filtered set of system noise data.

2. The method of claim 1, wherein the noise buffer is mapped into virtual address spaces of the plurality of parallel threads and an operating system kernel.

3. The method of claim 1, wherein the set of system noise data at least identifies a delay experienced by the at least one parallel thread corresponding to system noise during the first computation interval.

4. The method of claim 1, wherein the second computation interval is immediately sequent to the first computation interval.

5. The method of claim 1, wherein the storing comprises:

storing the filtered set of system noise data within one or more queues.

6. The method of claim 1, further comprising:

determining that at a first set parallel threads in the plurality of parallel threads has reached a synchronization barrier during the second computation interval prior to at least a second set of threads in the plurality of parallel threads; and

selecting, based on the determining, at least one parallel thread in the first set of parallel threads to perform the filtering.

7. The method of claim of claim 6, wherein the at least one parallel thread that has been selected comprises a second fastest computation time among the first set of parallel threads during at least the second computation interval.

8. The method of claim 1, further comprising:

reporting, during a third computation interval, the filtered set of system noise data, wherein the reporting comprises at least one of

storing the filtered set of system noise data in a local memory system; and

sending the filtered set of noise to at least one information processing system.

9. The method of claim 8, wherein the second computation interval is immediately subsequent to the first computation interval, and wherein the third computation interval is immediately subsequent to the second computation interval.

10. The method of claim 8, further comprising:

determining that at a first set parallel threads in the plurality of parallel threads has reached a synchronization barrier during the third computation interval prior to at least a second set of threads in the plurality of parallel threads; and

selecting, based on the determining, at least one parallel thread in the first set of parallel threads to perform the reporting.

11. The method of claim of claim 10, wherein the at least one parallel thread that has been selected comprises a fastest computation time among the first set of parallel threads during at least the third computation interval.

12. The method of claim 8, further comprising:

storing, during the second computation interval, at least one additional set of system noise data in the noise buffer, wherein the additional set of system noise data is detected during the second computation interval and is associated with at least one parallel thread in the plurality of parallel threads; and

filtering, during the third computation interval, the additional set of system noise data based on at least one filtering condition.

13. An information processing system for monitoring system noise in a parallel computing system, the information system comprising:

a memory;

a processor communicatively coupled to the memory; and

a system noise monitor communicatively coupled to the memory and the processor, wherein the system noise monitor is configured to perform a method comprising;

storing, during a first computation interval, at least one set of system noise data in a buffer, wherein the set of system noise data is detected during the first computation interval and is associated with at least one thread in a plurality of threads, and wherein each thread in the plurality of threads is a thread of a program;

filtering, during a second computation interval, the set of system noise data based on at least one filtering condition, the filtering creating a filtered set of system noise data; and

storing the filtered set of system noise data.

14. The information processing system of claim 13, wherein the method further comprises:

determining that at a first set parallel threads in the plurality of parallel threads has reached a synchronization barrier during the second computation interval prior to at least a second set of threads in the plurality of parallel threads; and

selecting, based on the determining, at least one parallel thread in the first set of parallel threads to perform the filtering.

15. The information processing system of claim of claim 14, wherein the at least one parallel thread that has been selected comprises a second fastest computation time among the first set of parallel threads during at least the second computation interval.

16. The information processing system of claim 13, wherein the method further comprises:

reporting, during a third computation interval, the filtered set of system noise data, wherein the reporting comprises at least one of

storing the filtered set of system noise data in a local memory system; and

sending the filtered set of noise to at least one information processing system.

17. The information processing system of claim 16, wherein the method further comprises:

determining that at a first set parallel threads in the plurality of parallel threads has reached a synchronization barrier during the third computation interval prior to at least a second set of threads in the plurality of parallel threads; and

selecting, based on the determining, at least one parallel thread in the first set of parallel threads to perform the reporting.

18. The information processing system of claim 17, wherein the at least one parallel thread that has been selected comprises a fastest computation time among the first set of parallel threads during at least the third computation interval.

19. A computer program product comprising a computer readable storage medium containing computer code that, when executed by a computer, implements a method for monitoring system noise in a parallel computing system, the method comprising:

storing, during a first computation interval, at least one set of system noise data in a buffer, wherein the set of system noise data is detected during the first computation interval and is associated with at least one thread in a plurality of threads, and wherein each thread in the plurality of threads is a thread of a program;

filtering, during a second computation interval, the set of system noise data based on at least one filtering condition, the filtering creating a filtered set of system noise data; and

storing the filtered set of system noise data.

20. The computer program product of claim 19, wherein the method further comprises:

determining that at a first set parallel threads in the plurality of parallel threads has reached a synchronization barrier during the second computation interval prior to at least a second set of threads in the plurality of parallel threads; and

selecting, based on the determining, at least one parallel thread in the first set of parallel threads to perform the filtering.

21. The computer program product of claim of claim 20, wherein the at least one parallel thread that has been selected comprises a second fastest computation time among the first set of parallel threads during at least the second computation interval.

22. The computer program product of claim 19, wherein the method further comprises:

reporting, during a third computation interval, the filtered set of system noise data, wherein the reporting comprises at least one of

storing the filtered set of system noise data in a local memory system; and

sending the filtered set of noise to at least one information processing system.

23. The computer program product of claim 22, wherein the method further comprises:

determining that at a first set parallel threads in the plurality of parallel threads has reached a synchronization barrier during the third computation interval prior to at least a second set of threads in the plurality of parallel threads; and

selecting, based on the determining, at least one parallel thread in the first set of parallel threads to perform the reporting.

24. The computer program product of claim 23, wherein the at least one parallel thread that has been selected comprises a fastest computation time among the first set of parallel threads during at least the third computation interval.

25. A method for monitoring system noise in a parallel computing system, the method comprising:

storing, during at least computation interval N, at least a first set of system noise data in a noise buffer, wherein the first set of system noise data is detected during the at least one computation interval and is associated with at least one parallel thread in a plurality of parallel threads, and wherein each thread in the plurality of parallel threads is a thread of a program;

filtering, during computation interval N, at least a second set of system noise data stored in the noise buffer, wherein the second set of system noise data was stored in the noise buffer during a computation interval N−

1, and where the second set of system noise data is associated with at least one parallel thread in the plurality of parallel threads; and

reporting, during computation interval N, at least a third set of system noise data, wherein the third set of system noise data was stored in the noise buffer during a computation interval N−

2, and where the third set of system noise data is associated with at least one parallel thread in the plurality of parallel threads.