[Performanceoptimization-wg] BBH benchmark results
schnetter at gmail.com
Sun Apr 29 12:35:03 CDT 2018
I ran benchmarks on Graham for Ian's BBH benchmark parameter files. Since
this measures OpenMP performance, I used a single node with 32 cores,
varying the number of OpenMP threads. The results are in the repo next to
the benchmark parameter files in a CSV file.
The result is: With a single OpenMP thread, as expected no difference. For
many (16) OpenMP threads, the new code is twice as fast. For intermediate
cases, the new code is up to 10x faster. This is clearly because the old
code behaves badly.
Erik Schnetter <schnetter at gmail.com>
More information about the performanceoptimization-wg