Standard Performance Evaluation Corporation
The SPEC MPI® 2007 benchmark suite is for evaluating MPI-parallel, floating point, compute intensive performance across a wide range of cluster and SMP hardware. This suite continues the SPEC tradition of giving users the most objective and representative benchmark suite for measuring and comparing high-performance computer systems.
The current release of the benchmark is V2.0.1.
The SPEC MPI® 2007 benchmark focuses on performance of compute intensive applications using the Message-Passing Interface (MPI), which means this benchmark emphasizes the performance of:
It is important to remember the contribution of all these components. Performance intentionally depends on more than just the processor; the benchmark is not intended to stress other computer components such as the operating system, graphics, or the I/O system.
This release (February 2010) adds the new SPEC MPI® L2007 suite to the SPEC MPI® 2007 benchmark suites. The new MPI® L2007 suite is designed for larger clusters, scaling up to 2048 ranks (cores). For an overview of the changes, see Changes in V2.0.
Results from V2.0 for the SPEC MPI® M2007 suite are comparable to those from V1.1 or V1.0. Results from the SPEC MPI® L2007 suite are not comparable to results from the SPEC MPI®M2007 suite and these results may continue to be submitted using V1.1 for the foreseeable future.
This maintenance release (October 2008) adds updates to benchmarks to improve portability and code correctness; improvements to report readability; clarifications and additions to the Run and Reporting Rules; documentation of the monitor hooks; and easier investigation of alternate sources or workloads. For a full list of updates, see Changes in V1.1. Existing licensees will receive a complimentary upgrade DVD.
Results from V1.0 are comparable to those from V1.1. As of January 1, 2009 all submissions must use version 1.1.
SPEC's HPG group recently voted to not allow the use of automatic parallelization compiler flags in the SPEC MPI® 2007 suite. This change effectively renders the benchmark as a pure MPI benchmark and is stated as such in newly-added section 1.1.4 of the run rules. Any reference to auto parallelization has been removed from the document and the reporting field "auto_parallel" in the results must always be set to "no" for a compliant result.
Fortunately, none of the results reviewed and published so far use the auto parallel option. Furthermore, please note that any "estimated" result, not reviewed by SPEC HPG, that makes use of the auto parallel option will no longer be compliant.
Journal articles, papers, and other publications related to the benchmark.