Errata for SPEC MPI2007
Last updated: 16 July 2007 blw
This document describes known problems in SPEC MPI2007. The latest version of this document may be found at
http://www.spec.org/mpi2007/Docs/errata.html.
- Problems may be added in the Open section from time to time.
- When problems are fixed, they are moved to the Closed section.
Notice to maintainers: When updating this document, please do not ever re-use problem numbers.
Open Problem Reports
4. ltest dataset may fail at 2048 ranks for 128.GAPgeofem
- Problem summary: When using Intel 11.1 compiler, SGI MPI on Intel
based hardware, a failure of the ltest dataset was seen on 128.GAPgeofem
when running at 2048 ranks.
- Details: The problem is being investigated, but has been shown to work
with other Intel 11.1 and other MPI combinations.
- Workaround: If you must use this software combination, you can use
ltrain rather than ltest for short tests at this rank count.
3. Windows Testing
- Problem summary: Testing of version 2.0 and 1.1 on Windows is very limited.
- Details: Very little testing on Windows has been attempted.
- Workaround: None at this time.
Closed Problem Reports
1. 126.lammps numerical sensitivity
- Problem summary: 126.lammps can be numerically sensitive
- Details: 126.lammps can get wrong answers or outright fail if the
numerics of the compiler are allowed to be liberal.
- Workaround: For the IBM compiler, you may need to use -qstrict to
get a good run.
2. 128.GAPgeofem large rank count
- Problem summary: Running 128.GAPgeofem at 1024 ranks caused a
run forever problem.
- Details: Same as summary.
- Workaround: The medium benchmark suite is claimed to support
up to 128 ranks. The benchmark has been successfully run up to 512
ranks. We have one example now where 1024 ranks hangs.
Do not attempt running at rank counts this high, the benchmark
may not run correctly.
Copyright © 2007-2010 Standard Performance Evaluation Corporation
All Rights Reserved