search

pmodels / mpich

Official MPICH Repository
http://www.mpich.org
Other
534 stars 280 forks source link

Comp_split and cart_create perfomance problems at scale #2333

Open mpichbot opened 7 years ago

mpichbot commented 7 years ago

Originally by gropp on 2016-04-10 16:30:26 -0500

One of my staff did a performance test on Vulcan, a BG/Q at LLNL, and noticed very poor performance at scale when 2 or 4 threads per core were used (and since oversubscription of cores is often important for performance, this is a common case). The test code is attached, as is the performance plot for comm_split.

mpichbot commented 7 years ago

Originally by gropp on 2016-04-10 16:31:46 -0500

Attachment added: CartCreateOnVulcan.png (11.3 KiB) CartCreateOnVulcan.png

mpichbot commented 7 years ago

Originally by gropp on 2016-04-10 16:31:55 -0500

Attachment added: SplitOnVulcan.png (11.5 KiB) SplitOnVulcan.png

mpichbot commented 7 years ago

Originally by gropp on 2016-04-10 16:32:03 -0500

Attachment added: testMPI.c (1.1 KiB)

mpichbot commented 7 years ago

Originally by robl on 2016-04-11 09:26:06 -0500

I suspect #1952 might help with this (Adam Moody contributed a bitonic sort for comm split). His patch got a little out of date.