Adaptive and dynamic filtering of threaded programs

• US 7,703,094 B2
• Filed: 12/30/2004
• Issued: 04/20/2010
• Est. Priority Date: 12/30/2004
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Armstrong, Douglas R, Shah, Sanjiv M, Mongkolsmai, Ekarat T
• Primary Examiner(s): An; Meng-Ai
• Assistant Examiner(s): KESSLER, GREGORY AARON
• Application Number:
  US11/026,687
• Publication Number:
  US 20060150162A1
• Time in Patent Office: 1,937 Days
• Field of Search: 718/100
• US Class Current: 718/100
• CPC Class Codes: G06F 11/3466 Performance evaluation by tracing or monitoring
• CPC Class Codes: G06F 11/3612 by runtime analysis performance monitoring G06F11/3466
• CPC Class Codes: G06F 2201/81 Threshold
• CPC Class Codes: G06F 2201/86 Event-based monitoring
• CPC Class Codes: G06F 2201/865 Monitoring of software

Abstract

A method and apparatus for adaptive and dynamic filtering of threaded programs. An embodiment of a method comprises analyzing the operation of a computer program, the computer program comprising a plurality of program threads; tracking overhead for the computer program; observing program events for the computer program; rationing overhead between program threads in inter-thread program events; and filtering program events based on a dynamic threshold.

Claims

1. A method comprising:

analyzing by a performance analyzer the operation of a computer application, the computer application comprising a first program thread and a second program thread, the performance analyzer including an adaptive filter;

recording overhead for the first program thread and the second program thread in a computer memory;

observing an inter-thread event comprising the transfer of a signal from the second program thread to the first program thread; and

adjusting the overhead for the first program thread using the adaptive filter if the first program thread waits a time period before receiving the signal from the second program thread, wherein adjusting the overhead for the first program thread includes;

determining whether the overhead of the second program thread minus the overhead of the first program thread is greater than the wait time period, and

upon determining that the overhead of the second program thread minus the overhead of the first program thread is greater than the wait time period, performing one of adding the wait time period to the overhead of the first program thread or ignoring the wait time period of the inter-thread event.

2. The method of claim 1, further comprising adding the overhead of the second program thread minus the overhead of the first program thread to the overhead of the first program thread if the overhead of the second program thread minus the overhead of the first program thread is not greater than the wait time period but is greater than zero.

3. The method of claim 1, further comprising making no change to the overhead of the first program thread if the overhead of the second program thread minus the overhead of the first program thread is not greater than zero.

4. The method of claim 1, further comprising establishing a threshold for analysis using a dynamic filter of the performance analyzer, wherein a second inter-thread event is not observed if the overhead resulting from the second inter-thread event is below the threshold.

5. A performance analyzer system comprising:

a dynamic memory to store a multithreaded program, the program comprising a first program thread and a second program threat

a processor to run the multithreaded program; and

a performance analyzer including;

an adaptive filter for the multithreaded program, the adaptive filter to determine the rationing of overhead between a plurality of program threads in inter-thread events, wherein the adaptive filter is to determine whether an inter-thread event between a first program thread and a second program thread, the first program thread waiting a time period before receiving a signal from the second program thread, may have occurred without the operation of the performance analyzer by determining whether the overhead of the second program thread minus the overhead of the first program thread is greater than the wait time period; and

a dynamic filter for the multithreaded program, the dynamic filter to modify a threshold for collection of data based on the performance of the program.

6. The performance analyzer system of claim 5, wherein the dynamic filter is to adjust the threshold based at least in part on a percentage of inter-thread events that are observed by the analyzer.

7. The performance analyzer system of claim 5, wherein the dynamic filter is to adjust the threshold based at least in part on a ratio of overhead time for the thread versus user time for the program.

8. A system comprising:

a dynamic memory to store a multithreaded program, the program comprising a first program thread and a second program thread;

a processor to run the multithreaded program; and

an analysis tool to analyze the performance of the multithreaded program as run by the processor, the analysis tool comprising a first component to track overhead for program threads and a second component to determine which cross-thread events are to be observed by the analysis tool;

wherein the first component determines the rationing of overhead time between the first program thread and the second program thread in a cross-thread event comprising the second program thread sending a signal to the first program thread, the first component determining an overhead element to ration to the first program thread if the first program thread waits for a time period before receiving the signal from the second program thread, the overhead element rationed to the first program thread being equal to the waiting period if the difference between the overhead of the second thread and the overhead of the first thread is greater than the length of the waiting time period.

9. The system of claim 8, wherein the overhead element rationed to the first program thread is the difference between the overhead of the second thread and the overhead of the first thread if the difference between the overhead of the second thread and the overhead of the first thread is not greater than the length of the waiting time period but is greater than zero.

10. The system of claim 8, wherein the overhead element rationed to the first program thread is zero if the difference between the overhead of the second thread and the overhead of the first thread is not greater than zero.

11. The system of claim 8, wherein the second component utilizes a threshold to determine which program cross-thread events are to be observed by the analysis tool.

12. The system of claim 11, wherein the second component dynamically changes the threshold based at least in part on the amount of overhead time incurred.

13. The system of claim 11, wherein the second component dynamically changes the threshold based at least in part on the number of cross-thread events observed.

14. A machine-readable medium having stored thereon data representing sequences of instructions that, when executed by a processor, cause the processor to perform operations comprising:

analyzing the operation of a computer program, the computer program comprising a plurality of program threads;

tracking overhead for the computer program;

observing inter-thread events for the computer program;

rationing overhead between program threads in inter-thread inter-thread events, wherein rationing overhead includes;

observing a first inter-thread event comprising the transfer of a signal from a second program thread to a first program thread; and

adjusting the overhead for the first program thread if the first program thread waits a time period before receiving the signal from the second program thread, wherein adjusting the overhead for the first program thread includes;

determining whether the overhead of the second program thread minus the overhead of the first program thread is greater than the wait time period, and

upon determining that the overhead of the second program thread minus the overhead of the first program thread is greater than the wait time period, performing one of adding the wait time period to the overhead of the first program thread or ignoring the wait time period of the inter-thread event; and

filtering program events based on a dynamic threshold.