Original URL: https://www.theregister.com/2007/08/03/spec_benchmark_mpi2007/

SPEC dishes out MPI test

How chatty is your supercomputer?

By Austin Modine

Posted in Systems, 3rd August 2007 23:15 GMT

The Standard Performance Evaluation Corporation (SPEC) has released a new benchmark suite that measures the performance of clusters running the much-loved - and much-despised - Message-Passing Interface protocols.

SPEC's new MPI2007 serves as a new brand of supercomputing benchmark. The fresh test suite helps distinguish the performance of CPUs, memory, communications interconnect, compilers and the shared file system in clusters. The benchmark ignores other components such as the OS, graphics or I/O system.

"SPEC saw the need within the research and science communities for a standardization benchmark suite based on real MPI applications," said SPEC chairman Kalyan Kumaran. "As with all of our benchmarks, SPEC MPI2007 benefits from SPEC's nearly two decades of experience in establishing source code and data sets, measurement tools, run rules and peer review that help ensure comparable and repeatable results."

The software test has two sets of metrics; one running from the CD without any tuning and a peak metric option, which allows for tuning of compilers and compiler flags to get the best performance. The measurements are normalized against a reference system to produce a "bigger-is-better" metric. The current reference is an eight-node cluster of Celestica A2210 systems with AMD's Opteron processors connected via GigE.

The software supports 32 and 64-bit architectures, and runs on Linux, Unix and Windows. A medium-sized data set in SPEC MPI2007 scales up to 512 processors. SPEC is currently working on larger data sets.

Thus far, AMD, HP, QLogic and SGI have hopped aboard the benchmark test. The results are available at SPEC's website here.

SPEC MPI2007 is available now for $800 on DVD. The new suite joins SPEC's other high performance computing benchmarks, HPC2002; for parallel and distributed architectures, and OMP2001; for shared-memory parallel processing. ®