[yt-svn] commit/yt-doc: ngoldbaum: Added a tip explaining how to assign work to the compute nodes rather than individual processors in a parallel_objects loop.
Bitbucket
commits-noreply at bitbucket.org
Wed Dec 14 11:36:29 PST 2011
1 new commit in yt-doc:
https://bitbucket.org/yt_analysis/yt-doc/changeset/6094f9188564/
changeset: 6094f9188564
user: ngoldbaum
date: 2011-12-14 20:33:32
summary: Added a tip explaining how to assign work to the compute nodes rather than individual processors in a parallel_objects loop.
affected #: 1 file
diff -r 5b9c86452b8307e33e90986e1ea63a8858deec8e -r 6094f91885640079feac14823e9e12d277c31926 source/advanced/parallel_computation.rst
--- a/source/advanced/parallel_computation.rst
+++ b/source/advanced/parallel_computation.rst
@@ -332,6 +332,15 @@
There will be a sweet spot between speed of run and the waiting time in
the job scheduler queue; it may be worth trying to find it.
+ * If you are using object-based parallelism but doing CPU-intensive computations
+ on each object, you may find that setting :py:data:'num_procs' equal to the
+ number of processors per compute node can lead to significant speedups.
+ By default, most mpi implimentations will assign tasks to processors on a
+ 'by-slot' basis, so this setting will tell yt to do computations on a single
+ object using only the processors on a single compute node. A nice application
+ for this type of parallelism is calculating a list of derived quantities for
+ a large number of simulation outputs.
+
* It is impossible to tune a parallel operation without understanding what's
going on. Read the documentation, look at the underlying code, or talk to
other yt users. Get informed!
Repository URL: https://bitbucket.org/yt_analysis/yt-doc/
--
This is a commit notification from bitbucket.org. You are receiving
this because you have the service enabled, addressing the recipient of
this email.
More information about the yt-svn
mailing list