Benchmarks and Profiling With Cactus Applications
Cactus applications have been widely used for benchmarking for machine procurement and testing, for understanding performance characteristics, and for profiling and optimization studies. The Cactus Team regularly run the benchmarks on this page with different architectures, compilers, MPI implementations etc. More details about publications and reports resulting from this work can be found on our publications page. Sasanka Madiraju’s master thesis presents the performance of Cactus benchmarks on different machines and architectures. A 2005 report to NSF describes why Cactus provides a good benchmarking platform. Also see the News: CactusADM is now part of the 2006 edition of the SPEC benchmark suite.
To read a description of the significance and profile of BSSN_PUGH, along with performance and scaling results with various architectures, see “A Scientific Application Benchmark using the Cactus Framework”
We currently maintain five standard benchmarking applications:
- Bench_BSSN_PUGH: A numerical relativity code using finite differences on an uniform structured grid. It uses the CactusEinstein infrastructure to evolve a vacuum spacetime, using the BSSN formulation of the Einstein equations. It employs finite differencing in space, an explicit time integration method. Download Bench_BSSN_PUGH.tar.gz
- Bench_BSSN_Carpet: A numerical relativity code using mesh refinement. It is similar to BSSN_PUGH, but instead of a uniform structured grid it uses the Carpet driver. Download Bench_BSSN_Carpet.tar.gz
- Bench_Whisky_Carpet: A relativistic hydrodynamics code using mesh refinement. It is similar to BSSN_Carpet, but it solves the relativistic Euler equations in addition to the Einstein equations. Download Bench_Whisky_Carpet.tar.gz
- BenchIO_FlexIO: An I/O benchmark using John Shalf’s FlexIO library. It measures the speed of writing simulation data to disk using the IOFlexIO I/O method. Download BenchIO_FlexIO.tar.gz
- BenchIO_HDF5: An I/O benchmark using the HDF5 library. It measures the speed of writing simulation data to disk using the HDF5 I/O method. Download BenchIO_HDF5.tar.gz
- Bench_ADM: historical benchmark that has been replaced by BSSN_PUGH. It uses a different numerical formulation of the equations of general relativity (ADM). It also forms part of the SPEC CPU2006.
Each tarball contains a Cactus source tree with just the thorns necessary to run the benchmark.
We have combined the above tarballs into one: Cactus.tar.gz. The same source code (but without the benchmarking parameter files and thorn lists) is available from our CVS repositories with the tag benchmark_20051007.
We also maintain a number of results on our old site. The following link has previous results from for a number of different benchmarks and target architectures.