R, Go & Other New Benchmarks Added

[phoronix]

Phoronix: R, Go & Other New Benchmarks Added

For those relying upon the Phoronix Test Suite for your open-source, automated benchmark needs on Linux, BSD, Solaris, and other operating systems, a number of new test profiles were recently made available...

http://www.phoronix.com/scan.php?page=news_item&px=R-Go-More-New-Benchmarks

[MetricT]

FYI, if you're going to use R as a benchmark, you *REALLY* need to make sure you have linked it against an optimized BLAS if your benchmark makes heavy use of matrices. The reference BLAS included with R is dog slow. Here's an R benchmark I ran on one of our Ivy Bridge Xeons comparing the stock BLAS to an optimized BLAS (in this case OpenBLAS). You can see several cases where an optimized BLAS is 20 times faster than the reference.

https://www.dropbox.com/s/nfl4glk7utiepvv/R-benchmark.pdf?dl=0

OpenBLAS, GotoBlas, ATLAS, Intel MKL, and AMD ACML are good BLAS to choose from. For extra fun, you can use nVidia's cuBLAS and offload matrix operations to the GPU for even more speedup.

If you don't replace the BLAS, you are effectively benchmarking the reference BLAS rather than the underlying hardware.

[Michael]

FYI, if you're going to use R as a benchmark, you *REALLY* need to make sure you have linked it against an optimized BLAS if your benchmark makes heavy use of matrices. The reference BLAS included with R is dog slow. Here's an R benchmark I ran on one of our Ivy Bridge Xeons comparing the stock BLAS to an optimized BLAS (in this case OpenBLAS). You can see several cases where an optimized BLAS is 20 times faster than the reference.

https://www.dropbox.com/s/nfl4glk7utiepvv/R-benchmark.pdf?dl=0

OpenBLAS, GotoBlas, ATLAS, Intel MKL, and AMD ACML are good BLAS to choose from. For extra fun, you can use nVidia's cuBLAS and offload matrix operations to the GPU for even more speedup.

If you don't replace the BLAS, you are effectively benchmarking the reference BLAS rather than the underlying hardware.

The R benchmark is relying upon the system's R build, which is what they intend for being able to compare distribution's performance, etc.