[petsc-dev] HPC benchmarking
bsmith at mcs.anl.gov
Fri Sep 16 14:20:41 CDT 2016
The absolute first thing to start with is the streams benchmark. You must run it yourself (or have it run by someone you trust) you cannot rely on what a vendor or someone on the web says about the performance. We provide an MPI based version in PETSc in src/benchmarks/streams when it is run you must also insure that all proper process and memory affinities are set properly (this is not always easy) http://www.mcs.anl.gov/petsc/documentation/faq.html#computers.
You'll want to know the streams numbers per core, per node and across the machine.
This is the foundation for understanding the machine, any additional benchmarks are just gravy that can only slightly change your understanding of the machine.
Note this advice is only for PETSc-type "PDE" simulations, other types of simulations may need very different benchmarks.
> On Sep 16, 2016, at 11:43 AM, Chris Kees <cekees at gmail.com> wrote:
> Is there a set of HPC benchmarks that the core PETSc developers recommend for evaluating architectures? Meaning to evaluating claims about performance to support rational acquisition decisions on new systems. I recall speaking briefly with Jed at a conference about a benchmark that was more representative of PDE calculations, maybe representing a Krylov iteration arising from an unstructured mesh or something like that. I'm talking about long term acquisition of real workhorse petascale systems not the cutting edge.
More information about the petsc-dev