10.1 Assessing PSE Application Performance

The parallel profiler measures a parallel application's performance by capturing the amount of time the program spends in different code regions or execution activities. This section provides a basic understanding of performance issues that arise when HPF programs are run in under PSE. This knowledge is useful to develop a strategy for effectively applying the parallel profiler.

10.1.1 PSE Program Execution

Four major factors have a direct impact on the performance of HPF applications running in under PSE:

• Configuration - Hardware and software components comprising the execution environment including the network interconnect
• System Resource Tuning - System parameters can be adjusted to optimize performance for particular applications, such as stack limit
• Partition and Communications Selection - Partitions and messaging protocols with different run-time characteristics influence performance significantly
• Program Implementation - Algorithms exploiting parallelism and minimizing communications are critical to PSE application performance

You can use the parallel profiler to analyze program design issues, determine optimal resource selection and identify system parameters to modify, as shown in Figure 10-1.

Figure 10-1 PSE Performance Factors

Refer to Part IV for information on PSE system configuration and its effect on performance.

10.1.2 Parallel Profiling Methods

pprof provides two methods to analyze parallel program execution:

• PC sampling

  A nonintrusive method that examines the program counter address at regular intervals. The profiler tallies the program counter value at each tick, providing a statistical picture of where the program spends its time. Statistical timing is based on the number of sample counts in each routine and is measured exclusively from the other routine calls.

• Interval profiling

  An intrusive method that performs source-level hierarchical timing. It is hierarchical in the sense that the time is recorded for the execution of a subroutine, including the execution time of other routines called. It collects interval timing (the elapsed time between the start and the end of an event) and count information on the occurrences of events. An event is a construct such as a routine or DO loop occurring at specific locations in the code. Each event is distinguished from other events by the trace pattern originating the event. A trace pattern can include a line number or file name specific to the event.

  Interval profiling also provides communication profiles related to array distribution in HPF programs. Time is measured in real time and timing for each event is measured inclusively with other events being called by it. Interval profiling statistics reflect inclusive time - they contain data for lower-level events called from within an event. For example, time spent in a DO loop includes the time spent for routine calls, other DO loops, and so on, occurring within the DO loop.

Time is measured using the Alpha process cycle counter allowing for nanosecond resolutions. If the range of the cycle counter is exceeded for a single event such as a very long input/output (I/O) operation, pprof detects this condition and reports a warning.

10.1.3 Parallel Profiling Data

Each profiling method provides a default report format. With pprof options, you can customize the reports to focus on particular events, types, or use different data display parameters.

Table 10-1 provides an overview of the different types of profiling data.

Table 10-1 parallel profiling Data Types