Method, apparatus, and article of manufacture for performance analysis using semantic knowledge

Abstract

Methods and systems consistent with this invention analyze the performance of a program executed in a data processing system. Such methods and systems assign a semantic to the performance of the program, and measure the level of performance of the program based on the semantic. As part of assigning a semantic, such methods and systems indicate a class of processing of which to measure performance, and may define a suctitude associated with the class. Such methods and systems define the class as a processing function that could contribute to the poor performance of the program. As part of measuring the level of performance, such methods and systems measure the suctitude of the indicated class during program execution.

Description

FIELD OF THE INVENTION

The present invention relates to the field of performance analysis, and more particularly to performance analysis using semantic knowledge.

BACKGROUND OF THE INVENTION

Computer systems execute programs that solve complex computational problems.

Preferably, the programs achieve high levels of performance, reduce wasted computer resources, and execute at peak speed. “Performance analysis” is the process of analyzing and understanding the execution characteristics of programs to identify impediments that prevent programs from running at peak speed, or their highest level of performance.

The amount of information required to completely characterize the execution of a program is massive, however, and it is therefore difficult or impossible to analyze all the data manually. Current automatic “performance analyzers” present performance data textually or graphically and direct the user's attention to patterns that may indicate a performance problem. These tools, however, lack an understanding of the meaning, or “semantic knowledge,” of the analyzed program, which limits their effectiveness in solving performance problems.

For example, performance analyzers generally attempt to identify algorithms that ineffectively use computer resources. To do this, conventional performance analyzers may identify parts of a program that take a long time to execute. This heuristic, however, may be deceptive. For instance, such an analyzer would identify a well-written algorithm as a poorly-performing algorithm simply because it unavoidably requires a lot of time to execute. Such an analyzer would also fail to identify poorly-performing algorithms because they do not take a long time to execute or because they are not central to the program. Without knowledge of the semantics of the programs, or how program components are supposed to run, an automatic performance analyzer cannot adequately determine whether a particular component of a program exhibits poor performance.

Performance analysis is also important in multiprocessing computer systems. A multiprocessing computer system comprises multiple processors in which different portions of a program execute in parallel in the different processors. Or, it is a system in which a program executes in parallel over multiple computers, each with a different processor. In such a computer system, resources may be wasted if processors are idle (i.e., not executing a program instruction) for any length of time. Thus, an automatic performance analyzer identifies algorithms that do not effectively divide tasks over the available processors, i.e., they have low “parallelism.” Conventional performance analyzers generally attempt to identify algorithms with low parallelism by indicating instances during program execution when one or more of the processors are idle. This may indicate when the program is not using the available processor resources as well as it could. Such a heuristic, however, may also identify instances when processors are expected to be idle, such as during the traversal of a linked list by a single processor. Further, even during the course of executing an extremely efficient program, the number of instances that one or more processors may be idle could be one billion or more. Conventional automated performance analyzers are incapable of distinguishing instances when the processors are expected to be idle from instances when they are not. Therefore, without knowledge of the semantics of the program, or how program components are supposed to run, automatic performance analyzers cannot adequately determine low parallelism portions of programs.

Thus, there is a need for performance analysis that identifies performance impediments based on an understanding of the meaning, or semantic knowledge, of the portions of the program being analyzed.

SUMMARY OF THE INVENTION

Methods and systems consistent with this invention analyze the performance of a program executed in a data processing system. Such methods and systems assign a semantic to the performance of the program, and measure the level of performance of the program based on the semantic. As part of assigning a semantic, such methods and systems indicate a class of processing of which to measure performance, and may define a suctitude, i.e. a degree of poor performance, associated with that class. Such methods and systems define the class as a processing function that could contribute to the poor performance of the program. As part of measuring the level of performance, such methods and systems measure the suctitude of the indicated class during program execution.

The summary and the following detailed description should not restrict the scope of the claimed invention. Both provide examples and explanations to enable others to practice the invention.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate an implementation of the invention and, together with the description, serve to explain the advantages and principles of the invention. In the drawings,

FIG. 1 depicts a data processing system suitable for use with methods and systems consistent with this invention;

FIGS. 2A, 2B, and 2C are exemplary displays, consistent with this invention, for providing output results of a performance analyzer; and

FIG. 3 is a flowchart of a process, consistent with this invention, for analyzing the performance of a program executed in the data processing system of FIG. 1.

DETAILED DESCRIPTION OF THE INVENTION

Overview

The following description of embodiments of this invention refers to the accompanying drawings. Where appropriate, the same reference numbers in different drawings refer to the same or similar elements.

Methods and systems consistent with this invention provide a performance analysis tool that identifies performance impediments based on an understanding of the meaning, or semantic knowledge, of the portions of the program being analyzed. Such methods and systems assign a semantic to the performance of the program, and then measure the level of performance of the program based on the semantic. The semantic may comprise a class of processing and a suctitude, i.e. a degree of poor performance, associated with the class. A class is anything that could contribute to the poor performance of a computer program. Use of semantic knowledge facilitates identification of impediments that prevent the program from executing a peak speed.

Implementation Details

FIG. 1 depicts a data processing system suitable for use with methods and systems consistent with this invention. Data processing system 100 comprises a computer 102 attached to a network 126. Computer 102 comprises a main memory 104, a first processor 116, a second processor 118, a secondary storage device 120, an input device 122, an output device 124, and a video display 114. Main memory 104 may comprise an operating system 112, an analyzed program 108, and a performance analyzer program 106. First processor 116 or second processor 118 execute analyzed program 108 and performance analyzer 109, but not necessarily at the same time.

Methods and systems consistent with this invention first gather data concerning the execution characteristics of program 108. The process of gathering information for performance analysis is called “instrumentation.” Instrumentation requires adding instructions to analyzed program 108 so that when it executes, these instructions generate data from which performance analyzer 106 derives performance information. For example, in “subprogram level instrumentation,” each subprogram is instrumented with a set of instructions that generate data reflecting calls to the subprogram. This allows, for example, the tracking of the number times the subprogram is called. Performance analyzer 106 may analyze the instrumentation data generated during execution of program 108 after program 106 is finished executing. Alternatively, performance analyzer 106 may analyze the data during execution of program 108. An example of instrumentation consistent with this invention is described in more detail below. Both the instrumentation and performance analyzer 106 may comprise “performance analysis.”

Methods and systems consistent with this invention analyze the performance of program 108, which is executed in data processing system 100. Such methods and systems assign a semantic to the performance of program 108, and measure the level of performance of the program based on the semantic. For instance, the semantic may take the form of a “class” and a “suctitude,” which is a degree of poor performance associated with the class. The class indicates the type of processing in terms that are meaningful in the context of application 108. For example, in a scientific application, classes may include “input,” “matrix multiplication,” and “output.” For multiprocessor computer systems, a class may be “idle processor,” or “stalled processor.” Another class may be “cache misses,” which occur when a cache memory for storing variable data is overwritten with other data. Other classes may be “sorting,” “searching,” “convolution,” or “decryption.” Essentially, a class may be anything that could contribute to the poor performance of a computer program, and not all classes consistent with this invention are listed here.

The suctitude indicates the degree to which the class constitutes a performance problem, where higher values may indicate larger problems. For example, if the class were “idle processor,” the suctitude may be defined as one. Thus, if processor 116 were idle, this would produce a calculated suctitude of one per unit of time. Ten idle processors would produce a calculated suctitude of 10 per unit of time. The suctitude of “stalled processor” may be 10, larger than the suctitude of an “idle processor,” because a stalled processor is more of a performance impediment than an idle processor. The unit of the elapsed time may be seconds, for example.

The user of performance analyzer 106 may dynamically indicate the classes that he or she believes is important for analysis, i.e., the classes he or she believes may inhibit good performance. Further, the user may dynamically define the suctitudes associated with the classes, reflecting his or her belief of the relative suctitude of the different classes. Alternatively, the classes and suctitudes are already indicated or defined in performance analyzer 106 by the software vendor.

Methods and systems consistent with this invention may divide classes into subclasses. For example, if the class were “stalled processor,” the subclasses could be (1) “stalled as a result of user,” or (2) “stalled as a result of system 100 activity.” Processor 116 may stall as a result of system 100 activity if, for instance, processor 116 is waiting for input/output to complete (or waiting for access to data), waiting for system resources to be assigned, or waiting for a dependent piece of the program to complete. Processor 116 may stall as a result of user activity, for instance, if processor 116 is waiting for the user to input information through input device 122.

Methods and systems consistent with this invention may also indicate a plurality of classes of which to measure performance, and define a suctitude associated with each class. Such methods and systems may also calculate the aggregate suctitude. The aggregate suctitude at any given time is the sum of the calculated suctitudes of all classes.

A class's defined suctitude may be a constant or it may be a function of other execution data. For example, an idle processor may have a constant suctitude per unit time, while other classes, such as “matrix multiplication,” may have a defined suctitude that is a function of the amount of work done and the time taken to do the work. In matrix multiplication, a square matrix M may have N rows and columns. The amount of time taken to square this matrix (matrix multiplication) may be proportional to the cube of the size of the matrix, or N³. Therefore, the suctitude may be a function defined by B=T−N³, where B is the suctitude and T is the total time it took for the matrix multiplication. If N were 10, and if the total amount of time T to square the matrix were 1000 units, the calculated suctitude B would be zero. If, on the other hand, the total amount of time T to square that matrix were 4000 units, then the suctitude B would be 3000. To the extent to which it takes more time than expected to perform the matrix multiplication, the higher the calculated suctitude.

The “matrix multiplication” class is used here to describe one example of performance instrumentation. When program 108 enters matrix multiplication subroutine, it records in secondary storage 120 (1) a start time when it entered the subroutine, and (2) the size of matrix M. When program 108 leaves matrix multiplication subroutine, it records (3) a leave time in secondary storage device 120. Performance analyzer 106 may analyze this data at a later time after program 108 finishes execution. Program 108 is instrumented to record all the data necessary for performance analyzer 106 to perform the semantic analysis described above. In this instance, performance analyzer 106 knows the size of matrix M and the amount of time it took perform the matrix multiplication and can calculate the suctitude. Because the suctitude may not be calculated during the execution of the program, the user may dynamically indicate and define the classes and suctitudes, as described above. The other classes in program 108 may similarly be instrumented, ensuring that the necessary data for the suctitudes to be calculated is recorded in secondary storage 120.

An example of an instrumented matrix multiplication subroutine follows:

   SUBROUTINE MATRIX_MULT(N,A,B,C)
   DOUBLE PRECISION A(N,N), B(N,N), C(N,N)
   INTEGER*8 START_NSEC, END_NSEC
   INTEGER*8 GETHRTIME !$PRAGMA C (GETHRTIME)
   START_NSEC = GETHRTIME( )
   DO 30,J = 1, N
   DO 20, L = 1, K
   TEMP = ALPHA*B(L,J)
   DO 10, I = 1, M
   C(I,J) = C(I,J) + TEMP*A(I,L)
10 CONTINUE
20 CONTINUE
30 CONTINUE
   END_NSEC = GETHRTIME( )
   CALL SAVE_START_STOP_AND_SIZE(START_NSEC,
   END_NSEC, N)
   END

This subroutine multiplies a matrix A by a matrix B and stores the result in a matrix C. Matrices A, B, and C are square of size N. Variable START_NSEC stores the time the subroutine starts, and variable END₁₃ NSEC sores the time the subroutine ends. Subroutine SAVE₁₃ START_STOP_AND_SIZE stores the variables START₁₃ NSEC, END₁₃ NSEC, and N to secondary storage device 120 so that performance analyzer may analyze this data at a later time. An example of the data stored to secondary storage device 120 follows:

• • 12003
  • 5261964
  • 10The first number 12003 is the start time of the matrix multiplication subroutine; the second number 5261964 is the end time of the matrix multiplication subroutine; and the last number 10 may be the size of the matrix multiplied.

Other classes whose suctitudes may be defined as a function of other execution data are “cache misses,” “input,” and “output.” For example, if the class were “cache misses,” five percent of memory requests ending up in cache misses may result in a calculated suctitude of 5 per unit time. If the class were “input” or “output,” the suctitude may be defined as a function of the amount of data input or output during a unit of time.

In a multiprocessor environment, two of the factors in the total performance of analyzed program 108 are (1) the amount of time spent waiting for another processor to finish, and (2) the amount of time spent executing code that cannot be executed in parallel. An example of code in a multiprocessor environment follows:

	CALL READ_DATA(N,A)	!	Single CPU only
	CALL PROCESS_DATA(N,A)	!	Multiple CPUs
	CALL WAIT_FOR_ALL_CPUS_TO_FINISH

Subroutine READ_DATA may be code that cannot be executed in parallel; subroutine PROCESS_DATA may efficiently execute in parallel processors; and subroutine WAIT_FOR₁₃ ALL_CPUS_TO_FINSH is executed while waiting for all the CPUs to finish executing code. Such a code may be instrumented by storing in secondary storage device 120 (1) the amount of time spent in READ_DATA, and (2) the amount of time spent in WAIT_FOR_ALL_CPUS_TO_FINSH, and then assigning a suctitude to each. READ_DATA may have a higher defined suctitude relative to WAIT_FOR_ALL_CPUS_TO_FINSH because single-CPU regions may be more of a performance impediment to parallel performance than synchronization.

Once the execution data has been gathered and suctitudes defined, performance analyzer 106 may use the calculated suctitude data to identify performance impediments in program 108. Methods and systems consistent with this invention display the calculated suctitude of a class as a function of time. In this case, the display may be a line graph with the suctitude on the ordinate, or Y-axis, and time on the abscissa, or X-axis. U.S. Pat. No. 6,434,714, entitled “Methods, Systems, and Articles of Manufacture for Analyzing Performance of Application Programs,” hereby incorporated by reference, describes ways of displaying performance analysis results. Performance analyzer 106 may display the calculated aggregate suctitude of all the classes as a function of time, or it may display the calculated suctitude of only one or a select few classes. Thus, the user may isolate the suctitude of a particular class.

Alternatively, methods and systems consistent with this invention display when the calculated suctitude of a class or group of classes exceeds a threshold. Or, performance analyzer 106 may indicate when during execution in the program the calculated suctitude reached a maximum. The calculated suctitude may be displayed as a, color, with different colors representing different numerical levels of calculated suctitude.

FIGS. 2A, 2B, and 2C are exemplary displays, consistent with this invention, for providing output results of performance analyzer 106 after analyzing the multiprocessor code example provided above. In FIGS. 2A, 2B, and 2C, the word “badness” is synonymous with suctitude. As shown in FIG. 2A, a user of performance analyzer 106 may define the READ_DATA subroutine suctitude to “5” in the “Set Badness” dialog box. The user may also define the suctitudes for the other routines in a similar fashion. For example, the suctitude of PROCESS_DATA may be defined as zero, and the suctitude of WAIT_FOR_ALL_CPUS_TO_FINISH may be defined as one. FIG. 2B shows the calculated suctitude with respect to time. Area 202 is the time period when subroutine READ_DATA executes; area 204 is the time period when PROCESS_DATA executes; and area 206 is the area when WAIT_FOR_ALL_CPUS_TO_FINISH executes. Area 202 is dark, showing a calculated suctitude of 5; area 204 is white, showing a suctitude of zero; and area 206 is light gray, showing a suctitude of one. FIG. 2C shows suctitude bookmarks 208, where the suctitude ranges between 1 and 5. Bookmarks 208 direct the user of performance analyzer 106 to particular areas of suctitude.

FIG. 3 is a flowchart of a process, consistent with this invention, for analyzing the performance of a program executed in data processing system 100. First, methods and systems consistent with this invention indicate a class of processing (step 302). Then such methods and system define a suctitude associated with the class indicated in step 302 (step 304). Steps 302 and 304 may be considered as assigning a semantic to the performance of a program. Such methods and systems then calculate the suctitude during an execution of the program (step 306). Calculating the suctitude in step 306 may be considered measuring the level of performance, which is then displayed as the result of the analysis (step 308). Process 300 may be implemented by program analyzer 106 that resides in main memory 104. Program analyzer may be stored on a computer-readable medium, such as secondary storage device 120. Therefore, the computer-readable medium may contain instructions for controlling computer system 100 to perform process 300 as a method of operating processors 116 or 118.

Systems consistent with this invention are applicable to all programs written in all computer programming languages, including Fortran 77, Fortran 95, Java, C, C++, and assembler for any given computer.

One skilled in the art appreciates that numerous variations to this system exist. For example, the performance data may be tabulated and displayed in any fashion. Although methods and systems consistent with this invention have been described with reference to a preferred embodiment thereof, those skilled in the art know various changes in form and detail which may be made without departing from the spirit and scope of this invention as defined in the appended claims and their full scope of equivalents.

Claims

1. A method for analyzing a performance of a program executed in a data processing system, the method comprising: assigning a semantic to the performance of the program; and measuring the performance of the program based on the semantic, wherein the performance is a function of an expected performance.

2. The method of claim 1, wherein assigning the semantic comprises indicating a class of processing of which to measure performance; and calculating a suctitude associated with the class during program execution.

3. The method of claim 2, wherein the suctitude indicates a performance problem associated with the class.

4. The method of claim 2, wherein the suctitude is a function of work done in a period of time and an expected work done in the period of time.

5. The method of claim 2, wherein the suctitude is a function of time.

6. The method of claim 2, wherein the suctitude is a function of execution data.

7. The method of claim 2, wherein the suctitude is a function of work done and expected work done.

8. The method of claim 2, further comprising: displaying a level of the suctitude of the class as a function of time.

9. The method of claim 2, further comprising: displaying when a level of the suctitude of the class exceeds a threshold.

10. The method of claim 2, wherein the class is one of input, matrix multiplication, output, idle processor, stalled processor, sorting, searching, convolution, decryption, and cache misses.

11. The method of claim 2, wherein the class of processing is a function of processing that contributes to the poor performance of the program.

12. The method of claim 2, wherein indicating a class comprises indicating a plurality of classes of processing of which to measure performance; and wherein defining the suctitude comprises defining a suctitude associated with each class.

13. The method of claim 2, further comprising: displaying an aggregate of suctitude of the plurality of classes as a function of time.

14. The method of claim 2, further comprising: displaying when a total level of the suctitude of the plurality of classes exceeds a threshold.

15. The method of claim 2, further comprising: instrumenting the program to allow the measuring of the suctitude for the class.

16. An apparatus for analyzing performance of a program, the system comprising: a processor configured to perform assigning a semantic to the performance of the program; and measuring the performance of the program based on the semantic, wherein the performance is a function of an expected performance.

17. The apparatus of claim 16, wherein assigning the semantic comprises indicating a class of processing of which to measure performance; and calculating a suctitude associated with the class during program execution.

18. The apparatus of claim 17, wherein the suctitude indicates a performance problem associated with the class.

19. The apparatus of claim 17, wherein the suctitude is a function of work done in a period of time and an expected work done in the period of time.

20. The apparatus of claim 17, wherein the suctitude is a function of time.

21. The apparatus of claim 17, wherein the suctitude is a function of execution data.

22. The apparatus of claim 17, wherein the suctitude is a function of work done and expected work done.

23. The apparatus of claim 17, further comprising: a display to display a level of the suctitude of the class as a function of time.

24. The apparatus of claim 17, further comprising: a display to display when a level of the suctitude of the class exceeds a threshold.

25. The apparatus of claim 17, wherein the class is one of input, matrix multiplication, output, idle processor, stalled processor, sorting, searching, convolution, decryption, and cache misses.

26. The apparatus of claim 17, wherein the class of processing is a function of processing that contributes to the poor performance of the program.

27. The apparatus of claim 17, wherein the processor is further configured to indicate a plurality of classes of processing of which to measure performance; and define a suctitude associated with each of the plurality classes.

28. The apparatus of claim 17, further comprising: a display to display the total level of the suctitude of the plurality of classes as a function of time.

29. The apparatus of claim 17, further comprising: a display to display when a total level of the suctitude of the plurality of classes exceeds a threshold.

30. The apparatus of claim 17, wherein the program comprises instrumented lines of code to allow the measuring of the suctitude for the class.

31. A computer-readable medium comprising instructions for controlling a computer system to perform a method of operating a processor, the method comprising: assigning a semantic to the performance of the program; and measuring the performance of the program based on the semantic, wherein the performance is a function of an expected performance.

32. The computer-readable medium of claim 31, wherein assigning a semantic comprises indicating a class of processing of which to measure performance; calculating a suctitude associated with the class during program execution.

33. The computer-readable medium of claim 32, wherein the suctitude indicates a performance problem associated with the class.

34. The computer-readable medium of claim 32, wherein the suctitude is a function of work done in a period of time and an expected work done in the period of time.

35. The computer-readable medium of claim 32, wherein the suctitude is a function of time.

36. The computer-readable medium of claim 32, wherein the suctitude is a function of execution data.

37. The computer-readable medium of claim 32, wherein the suctitude is a function of work done and expected work done.

38. The computer-readable medium of claim 32, wherein the method further comprises: displaying a level of the suctitude of the class as a function of time.

39. The computer-readable medium of claim 32, wherein the method further comprises: displaying when a level of the suctitude of the class exceeds a threshold.

40. The computer-readable medium of claim 32, wherein the class is one of input, matrix multiplication, output, idle processor, stalled processor, sorting, searching, convolution, decryption, and cache misses.

41. The computer-readable medium of claim 32, wherein the class of processing is a function of processing that contributes to the poor performance of the program.

42. The computer-readable medium of claim 32, wherein indicating a class comprises indicating a plurality of classes of processing of which to measure performance; and wherein defining the suctitude comprises defining a suctitude associated with each class.

43. The computer-readable medium of claim 32, wherein the method further comprises: displaying an aggregate of the suctitude of the plurality of classes as a function of time.

44. The computer-readable medium of claim 32, wherein the method further comprises: displaying when a total level of the suctitude of the plurality of classes exceeds a threshold.

45. The computer-readable medium of claim 32, wherein the method further comprises: instrumenting the program to allow the measuring of the suctitude for the class.