[R-sig-hpc] parallel and openblas

[Chris Paciorek]

This message is a follow-up to a thread from April 2012 that started
with this message:
https://stat.ethz.ch/pipermail/r-sig-hpc/2012-April/001339.html.

The issue in the thread was that OpenBLAS was causing the affinity of
R jobs (and child processes created within a given job) to be set to core 0 on
multi-core machines, with the result that child processes  created
through functionality in the parallel and multicore packages were all
being placed on the same core.

We've also experienced this issue on Linux machines running Ubuntu
12.04 with R 2.15.0 and 2.15.1. This has occurred on a variety of
chips (the new AMD Bulldozer (AMD Opteron 6272), Intel Core i7, and
AMD Opteron 8220).

Usually the first R job started on a machine had its affinity set.
Subsequent R jobs often, but not always, had their affinity set. We
solved the issue by setting NO_AFFINITY=1 when building OpenBLAS, as
suggested by Steve Weston in that thread (thanks, Steve!).

It didn't appear that anyone on the R side of things had told the
OpenBLAS community about this issue, so I posted to the OpenBLAS
mailing list and got this response from Xianyi Zhang,
http://list.rdcps.ac.cn/pipermail/openblas/2012-July/000115.html, in
which he states that it is not a bug but does cause the conflict with
R.  As I commented in response to his message, it does seem strange
that the affinity is sometimes set and sometimes not set in a given
session, suggesting some strange dependence on the state of the
system.

As far as next steps on the R side, the R installation manual
discusses GotoBLAS and OpenBLAS. As a suggestion to Simon or other
core team members who read r-sig-hpc, I think it would be worth
mention in the OpenBLAS section of the installation manual
that one should set NO_AFFINITY=1 when building OpenBLAS.

Chris

----------------------------------------------------------------------------------------------
Chris Paciorek

Statistical Computing Consultant, Lecturer, Assistant Researcher

Office: 495 Evans Hall                      Email: paciorek at stat.berkeley.edu
Mailing Address:                            Voice: 510-842-6670
Department of Statistics                    Fax:   510-642-7892
367 Evans Hall                              Skype: cjpaciorek
University of California, Berkeley          WWW: www.stat.berkeley.edu/~paciorek
Berkeley, CA 94720 USA                      Permanent forward:
paciorek at alumni.cmu.edu