Skip to Content.
Sympa Menu

lapackers - [lapackers] Reproducible BLAS

Subject: lapackers

List archive

[lapackers] Reproducible BLAS


Chronological Thread 
  • From: James Demmel <demmel@cs.berkeley.edu>
  • To: lapackers <lapackers@cs.berkeley.edu>, lapackers <lapackers@cs.utk.edu>, "scalapackers@cs.utk.edu" <scalapackers@cs.utk.edu>
  • Subject: [lapackers] Reproducible BLAS
  • Date: Mon, 13 Sep 2010 06:23:03 -0700

We are having a discussion in the ParLab about the costs and benefits
of guaranteeing "reproducible results" when running in parallel, i.e. you get
bitwise identical results if you type a.out twice in a row. Does anyone know
if any vendor's BLAS library has a reproducibility flag that guarantees
this, perhaps under some extra conditions, like constant number of
processors, or fixed data layouts (either of which could in principle
vary from run to run, depending on how dynamically data and processors
are allocated).
Thanks,
Jim



  • [lapackers] Reproducible BLAS, James Demmel, 09/13/2010

Archive powered by MHonArc 2.6.16.

Top of Page