Benchmarks for evaluating the performance of sparse linear algebra implementations. The code in this repository is taken from open source benchmark suites, real world chemical simulation software, and simulated workloads from other problem domains.

Results from these benchmarks appear in the following paper:

• Automatically Harnessing Sparse Acceleration (CC 2020)

Building

The benchmarks in this package must be built and run separately using their own build systems.

git clone https://github.com/Baltoli/sparsity.git
cd sparsity

LD_LIBRARY_PATH needs to be able to find the installed libspmv libraries (see below), as well as any required by the different platforms, both at compile and run time.

libspmv

To build the shared library linear algebra implementations:

cd libspmv
make {native,mkl,gpu,opencl,clgpu}.so

Each implementation requires different libraries to be installed. These dependencies are:

• native: The sequential C implementation has no dependencies and should build on any system with a C compiler to act as a performance baseline.
• mkl: Requires an Intel CPU with MKL installed. The environment variable $MKLROOT should point to the installation of MKL, and can be defined by running the shell scripts supplied with MKL itself.
• gpu: Requires a CUDA-compatible Nvidia GPU and CUDA 9.1 installed, with installation root $CUDA_ROOT.
• opencl: OpenCL version that uses the first available OpenCL device on your platform, with OpenCL installed in $CL_ROOT. On our test platform, the first available device is the integrated GPU.
• clgpu: As for OpenCL, but using the second device - this corresponds to the external GPU on our test platform.

To install a built shared library:

SPMV_ROOT=prefix make platform={native,mkl,gpu,opencl,clgpu} install

The installed library will be named libplatform-spmv.so.

NPB

We use the conjugate gradient benchmark from the NAS benchmark suite. To build and run it:

cd NPB3.3.1
mkdir bin
make SPMV_VERSION=platform F77=gfortran CLASS=C CG
./bin/cg.C.x

The relevant benchmark results are the MOp/s and seconds taken to run. F77 can be set to whatever fortran compiler is most appropriate.

SparseBench

This application implements sparse matrix benchmarks for various different conjugate gradient variants. We use the general, unsymmetric version.

The first step is to generate suitably large matrices:

./big_gen.py --filename crsmat170u --size 170 write