next up previous
Next: Memory Profiling Up: Measurement Previous: High Resolution Timers

Multiple Metric Measurement

Tracing and profiling use a single metric for performance measurements by default. During profiling, TAU measures inclusive and exclusive time, the number of calls, and the child calls for each routine. However, there are cases where the user may want to measure more than one metric. In addition to the wallclock time, counts of hardware performance metrics and message sizes may be useful. TAU supports access to hardware performance counters using the PAPI interface [19]. The user can configure TAU using the -MULTIPLECOUNTERS option and set environment variables COUNTER1-N to specify the nature of performance metrics that should be recorded. The user can set these counters by choosing from preset events provided by PAPI e.g., PAPI_FP_INS for floating point instructions as well as native events that are processor specific e.g., PAPI_NATIVE_event name.

Figure 1: Performance data for memory headroom analysis

Figure 2: Memory headroom available on node 0

Scott Biersdorff 2006-05-05