3.7.1. PILS (Parallel ImbaLance Simulator)ΒΆ

PILS is an MPI+OpenMP/OmpSs synthetic benchmark that measures the execution time of imbalanced MPI ranks.

Usage:

./mpi_ompss_pils <loads-file> <parallel-grain> <loops> <task_size>
    loads-file:     file with load balance (number of tasks per iteration) per process, [100, 250] if /dev/null
    parallel-grain: parallelism grain, factor between 0..1 to apply sub-blocking techniques
    loops:          number of execution loops
    task_size:      factor to increase task size

Goals of this exercise

  • Run the instrumented version of PILS and generate a Paraver trace.

    • Analyse the load imbalance between MPI ranks.

  • Enable DLB and compare both executions.

    • Observe the dynamic thread creation when other processes suffer load imbalance.
    • Analyse the load imbalance of the new execution. Does it improve?

  • Enable DLB MPI interception and trace again. Analyse the new trace.

  • Run the multirun.sh script and compare the execution performance with and without DLB.

  • Modify the inputs of PILS to reduce load imbalance and see when DLB stops improving performance.