Benchmarking

From SC07 Student Competition

In your team's home directory on bobsced, there is a file named IMB_3.0.tar.gz. This tar file contains benchmarking utilities for testing MPI performance across your cluster.

The task at hand is easily described: Compile the IMB suite and obtain MPI benchmarks for the bobsced cluster. There are a few twists, however.

Problem Description

Your task is to build the Intel MPI Benchmarks from source. In order to carry out this task, you will need to use make -f. In addition, you will need to control your shell environment variables, and to possibly (likely) modify the way your shell defaults to these settings through your ~/.profile file.

Caveats

• The IMB suite 'prefers' mpich or Intel's mpi. You may want to tune the environmental parameters
  • PATH
  • LD_LIBRARY_PATH
  • P4_GLOBMEMSIZE
• If you are looking for alternative MPI compilers, you might find running the command "locate mpich" and similar commands very illuminating.

We know this problem works using MPICH on the bobsced cluster, so try your best and you'll certainly push it('s) through(put).

Analysis

bobsced consists of a total of sixteen quad-core Xeon systems. At the minimum, examine 2-core and 4-core IMB performance for the full suite of benchmarks. Additional cores won't add much to the analysis of your solution to this task, so please don't go overboard and swamp the cluster (which would be detrimental to your score).