lkml.org 
[lkml]   [2013]   [Oct]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectBUG: mm, numa: test segfaults, only when NUMA balancing is on
Hi guys,

I ran into a bug a week or so ago, that I believe has something to do
with NUMA balancing, but I'm having a tough time tracking down exactly
what is causing it. When running with the following configuration
options set:

CONFIG_ARCH_SUPPORTS_NUMA_BALANCING=y
CONFIG_NUMA_BALANCING_DEFAULT_ENABLED=y
CONFIG_NUMA_BALANCING=y
# CONFIG_HUGETLBFS is not set
# CONFIG_HUGETLB_PAGE is not set

I get intermittent segfaults when running the memscale test that we've
been using to test some of the THP changes. Here's a link to the test:

ftp://shell.sgi.com/collect/memscale/

I typically run the test with a line similar to this:

./thp_memscale -C 0 -m 0 -c <cores> -b <memory>

Where <cores> is the number of cores to spawn threads on, and <memory>
is the amount of memory to reserve from each core. The <memory> field
can accept values like 512m or 1g, etc. I typically run 256 cores and
512m, though I think the problem should be reproducable on anything with
128+ cores.

The test never seems to have any problems when running with hugetlbfs
on and NUMA balancing off, but it segfaults every once in a while with
the config options above. It seems to occur more frequently, the more
cores you run on. It segfaults on about 50% of the runs at 256 cores,
and on almost every run at 512 cores. The fewest number of cores I've
seen a segfault on has been 128, though it seems to be rare on this many
cores.

At this point, I'm not familiar enough with NUMA balancing code to know
what could be causing this, and we don't typically run with NUMA
balancing on, so I don't see this in my everyday testing, but I felt
that it was definitely worth bringing up.

If anybody has any ideas of where I could poke around to find a
solution, please let me know.

- Alex


\
 
 \ /
  Last update: 2013-10-16 18:21    [W:0.080 / U:2.752 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site