Known Problems - SPEChpc 2021

Introduction

This document describes known problems in SPEChpc 2021.

$Id$	Latest: www.spec.org/hpg/hpc2021/Docs/

Do you have a problem with SPEChpc?
Try these, in this order:

Update your copy with runhpc --update.
Check the latest version of Frequently Asked Questions.
Check the latest version of Known Problems.
Check the src.alt list. [What is a src.alt?]
Contact SPEC.

SPEChpc 2021 was tested on a variety of systems prior to release, including x86, Power, and ARM Linux systems. Nevertheless, as with any software product, problems may be found too late in the testing process to be fixed in the release, or may be discovered after release.

If you discover problems that are not listed here or in the FAQ, please report them to SPEC.

About the format of this document:

Problems may be added in the Open section from time to time. In each case, a workaround is given, if one is known.

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.

Known Limitations

Benchmark	Description	Work around
LBM	LBM may fail to validate when using larger ranks counts, usually 2-4x that of the upper tested limit for each suite.	Use fewer ranks or use a suite intended for larger rank counts.
LBM	v1.1 added host parallelization for the initialization loop in OpenMP with Target Offload and OpenACC. OpenACC uses the "self" clause which is not supported by all compilers.	Add the define flag "-DSPEC_OPENACC_NO_SELF" to fallback to using an OpenMP "parallel loop" for OpenACC.
miniSweep	miniSweep may not validate with odd ranks counts (except 1 or 3).	Use even rank counts
SOMA	SOMA uses array reductions for OpenMP and OpenACC when the compiler claims support for the standard where array reductions were added. OpenMP=>201511, OpenACC=>201811. However some compilers do not actually include support for array reductions.	Add the define flag -DSPEC_NO_VAR_ARRAY_REDUCE to use an alternate source version that does not use array reductions. Note that since the alternate version is run serially on the host, it's expected to be less performant.

Benchmark

Description

Work around

LBM

LBM may fail to validate when using larger ranks counts, usually 2-4x that of the upper tested limit for each suite.

Use fewer ranks or use a suite intended for larger rank counts.

LBM

v1.1 added host parallelization for the initialization loop in OpenMP with Target Offload and OpenACC. OpenACC uses the "self" clause which is not supported by all compilers.

Add the define flag "-DSPEC_OPENACC_NO_SELF" to fallback to using an OpenMP "parallel loop" for OpenACC.

miniSweep

miniSweep may not validate with odd ranks counts (except 1 or 3).

Use even rank counts

SOMA

SOMA uses array reductions for OpenMP and OpenACC when the compiler claims support for the standard where array reductions were added. OpenMP=>201511, OpenACC=>201811. However some compilers do not actually include support for array reductions.

Add the define flag -DSPEC_NO_VAR_ARRAY_REDUCE to use an alternate source version that does not use array reductions. Note that since the alternate version is run serially on the host, it's expected to be less performant.

Open Problem Reports

Please see the latest on-line version of this document for information on open problems and suggested workarounds.

Benchmark	Description	Work around
No known issues.

Benchmark

Description

Work around

No known issues.

Closed Problem Reports

Benchmark	Description	Solution
miniSweep	miniSweep OpenMP may intermittently hang for certain compilers.	Fixed in v1.1. Removed the "depend" clause which was found to be the source of the hang.

Benchmark

Description

Solution

miniSweep

miniSweep OpenMP may intermittently hang for certain compilers.

Fixed in v1.1. Removed the "depend" clause which was found to be the source of the hang.

SPEChpc™ 2021 Known Problems

Introduction

Known Limitations

Open Problem Reports

Closed Problem Reports