Skip to content

Benchmarks for experimentation of performance of features of OpenMP in the SOLLVE project.

Notifications You must be signed in to change notification settings

SOLLVE/benchmarks

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Repository for Performance Benchmarks for OpenMP Features

This is a repository for benchmarks for performance experimentation of features of OpenMP 5.0 for United States Department of Energy's Exascale Computing Project's (ECP's) OpenMP project called SOLLVE. See sollve.github.io for more information on the SOLLVE project.

The benchmarks are based on SOLLVE's efforts to undertand performance of OpenMP features, particularly OpenMP offload features. These benchmarks have been experimented on various systems as shown in the table below.

System LLVM 11 - Summit LLVM 11 - Cori-GPU LLVM 11 - Theta Spock
Stencil - 1D tested, works tested, works tested, works
Stencil - 2D tested, works tested, works tested, works
MatMul tested, works tested, works tested, works
Load Imbalanced MatMul tested, works tested tested, works
Dot Product tested, works tested tested, works
Stream tested
Square Rooted Vec Prod tested, works tested, work tested, works
System LLVM 12 - Summit LLVM 12 - Cori-GPU LLVM 12 - Theta Spock
Stencil - 1D tested, works tested, works tested, works
Stencil - 2D tested, works tested, works tested, works
MatMul tested, works tested, works tested, works
Dot Product tested, works tested tested, works
Stream tested
Square Rooted Vec Prod tested, works tested, work tested, works
gauss-seidel tested, works tested, works tested, works
lu factorization tested, works tested, works tested, works tested, works
spmv tested, works tested, works tested, works tested, works
quicksort tested, works tested, works tested, works tested, works
barnes-hut tested, works tested, works tested, works tested, works
Load Imbalanced MatMul tested, works tested tested, works
System LLVM 12 - Summit LLVM 12 - Cori-GPU LLVM 12 - Theta Spock
Stencil - 1D tested, works tested, works tested, works tested, works
Stencil - 2D tested, works tested, works tested, works tested, works
MatMul tested, works tested, works tested, works tested, works
Dot Product tested, works tested tested, works tested, works
Stream tested
Square Rooted Vec Prod tested, works tested, works tested, works
gauss-seidel tested, works tested, works tested, works
lu factorization tested, works tested, works tested, works tested, works
spmv tested, works tested, works tested, works tested, works
quicksort tested, works tested, works tested, works tested, works
barnes-hut tested, works tested, works tested, works tested, works
Load Imbalanced MatMul tested, works tested tested, works tested, works

Each benchmark uses a subset of OpenMP offload features. The stencil, matmul, dot product, square rooted vec prod use a target offload construct. The load imbalanced matrix multiplication uses a device clause with the target construct. The stream benchmark uses an experimental features of target spread for multi-GPUs along with the target construct and devices clause.

About

Benchmarks for experimentation of performance of features of OpenMP in the SOLLVE project.

Resources

Stars

Watchers

Forks

Packages

No packages published