Statistical sampling methods


Event-based methods

GUI mode

To invoke CXperf in GUI mode, type cxperf at your UNIX prompt. Navigate through the profiling process using the Previous and Next buttons at the bottom of each page. You may be automatically moved between pages as the profiling process progresses; you may be prevented from moving to a page under certain conditions. By guiding you through the process in this way, profiling with CXperf becomes a straightforward and intuitive process.

CXperf provides graphical analysis of performance data through the GUI. The GUI provides:

• Mouse-driven selection of region types to profile and metrics to collect
• Intuitive, step-by-step guidance through the compile, instrument, execute, and analyze steps
• Summary Profile (2D) and Parallel Profile (3D) graphs to analyze performance data
• Call Graphs with point-and-click navigation
• Source code correlation when you click on a bar of the Summary or Parallel Profile graph or on a node in the Call Graph
• Source code display facility with source code annotations indicating regions profiled.
• Text performance report functionality on the Analysis Page
• PDF analysis for files created on a different architecture
• Multiple Analysis Page comparisons which allow you to:
  • Compare and contrast profiling data for different metrics
  • View data from multiple PDFs simultaneously.

For detailed information on profiling a program in GUI mode, see the HP CXperf's User's Guide.


Line mode

Line mode is a character based, command line interface for CXperf. To use line mode, specify the -nw option when you invoke CXperf from the UNIX prompt. When you start CXperf in line mode with the name of a PDF, use the set pdf and analyze commands to access performance data for multiple PDFs, including PDFs created on different architectures.

Line mode presents performance data in Text reports only. However, after you collect profiling data in line mode, the resulting PDF can be analyzed in GUI mode.

Detailed information on profiling a program in line mode is presented below.

Batch mode

Batch mode allows you to run CXperf by incorporating it in a script or text file. You make use of CXperf's tty commands to profile applications in batch mode.

To use batch mode, provide a command file using the -x option on the command line, or invoke CXperf from a shell script, or both - provide the command file within a shell script. You can redirect input, output, and standard error to and from files.

For detailed information on profiling a program in batch mode, see the HP CXperf's User's Guide.

Compiling

The first step in the performance analysis process is compilation. CXperf does not actually aid you in compiling, but provides instructions - in GUI mode - for compiling programs using HP Fortran 90, ANSI C++, ANSI C, and HP Parallel 32-bit Fortran 77 compilers. To see the Compilation Page with the instructions you must start CXperf in GUI mode. Enter cxperf at your UNIX prompt:

% cxperf

The Compilation Page appears, displaying instructions for compilation.

Table 1 lists instructions for compiling your program.

Step 1.

Read the compile instructions in Table 1.

Step 2.

Decide which compile preference you need. For this example, compile and link in a single step.

Step 3.

Compile and link in a single step to analyze routines and loops. For example, using the ANSI C compiler enter: % /opt/ansi/bin/c89 +pal +O3 myprogram When compilation completes, you have an executable file, a.out, ready for profiling with CXperf. Your profiling options are determined by the compilation you just performed: +pal compiles the program for routine and loop level profiling. +O3 specifies a compiler optimization level that supports profiling routines and loops.

Table 1 describes compiling command syntax to use for programs compiled with HP ANSI C (c89), ANSI C++ (aCC), Fortran 90 (f90), and Fortran 77 (f77) compilers.



Refer to Preparing programs to profile for more details about compiling source, object, and library files for CXperf.


Instrumenting

To instrument your executable file, first invoke CXperf.

To start CXperf and instrument a.out in line mode, follow this procedure:

Step 1.

Invoke CXperf with the name of your executable file and the -nw (no windows) option. % /opt/cxperf/bin/cxperf -nw a.out Convex Performance Analyzer Type 'help' for help. Reading executable a.out... Selecting profile a.out.pdf... (CXperf) As shown in the output for the command above, CXperf displays the name of the executable file to be profiled and the name of the PDF that the performance data is written to. By default, the PDF is named executable.pdf.

Note:

Use the CXPERF environment variable to specify command line options for starting CXperf. For example, % setenv CXPERF `-nw -pid -w' forces CXperf to start in line mode (-nw). The -pid option specifies that CXperf add the process ID number of the process you are profiling to the name of the PDF it creates. The -w option suppresses warning messages issued by CXperf.

Step 2.

Select regions to profile with a form of the select command. The select command syntax is as follows: select [ routine | loop ] all where routine   Selects routines to profile. loop   Selects loops to profile. all   Instructs CXperf to select all routines in your program for profiling. In this example enter: (CXperf)   select all Because a.out is compiled with the +pal option, for this example, routines and loops are available for instrumentation. The all parameter instructs CXperf to select all routines in your program for profiling. Both routines and loops are selected for profiling because you did not use the [ routine | loop ] parameter to specify only one region type. In line mode, if you do not use select to select one or more source code regions for profiling, CXperf does not collect any metrics. For detailed information on using the select command, see Selecting routines and loops in the HP CXperf's User's Guide. If you have a large program, do not select all routines and all loops to profile in a single session, because the more region types and metrics you select, the slower your code executes. This time delay is referred to as profiling intrusion. For a detailed discussion of profiling intrusion, see Profiling strategy in the HP CXperf's User's Guide.

Step 3.

Select metrics to collect with the collect and set events commands. (CXperf)   collect cpu wall_clock call_graph events (CXperf)   set events process collect instructs CXperf to collect CPU time Wall Clock time Call Graph events (Specifies collecting one metric set available on the current architecture. Memory, Process, Data Cache Utilization, and Data and Instruction TLB events are possibilities.) Use set events immediately after collect events to specify which events to collect. For this example, set events process instructs CXperf to collect Process events. The type and number of metrics available differ according to machine architecture. For details, refer to Introducing metrics in the HP CXperf's User's Guide.

Executing

This section describes how to execute your program under CXperf control in line mode.

Note:

As an alternative to executing your program, you can write the instrumentation selections you just made to the executable file, using the save executable command. You can then exit CXperf and run the executable file to generate a Performance Data File (PDF). For further details, refer to Preinstrumenting in line mode in the HP CXperf's User's Guide.

Step 1.

Run your program using the run command. (CXperf) run The run command syntax is: run [ argument ... ] [ i/o_redirection ] where argument   Specifies any number of command line arguments to the program you are profiling. Separate multiple arguments with spaces. io_redirection   Redirects the standard input, output, or error from or to the specified file when you use one of the redirection operators (<, >, >>, >&, >>&).

Step 2.

Wait for the program to complete running. Your program runs to completion unless you press CTRL-C to pause it. For the most accurate results, do not pause your program during profiling. When you pause a program, use continue to resume execution, or stop to terminate the program.

Analyzing

Send comments/questions to: help-hpc@uky.edu
Last modified: August 02 2002 09:57:54.