Big slices vs Linux THP defrag

[Albert Strasheim]

Hello all

We ran into an interesting problem today where a Go program making a few 1 GB slices seems to spend forever allocating this memory (Linux, amd64).

A look a perf top showed a very busy compaction_alloc function.

The problem seems to be that the big allocation causes a Transparent Hugepages defrag operation which takes forever.

Disabling THP fixed the problem:

echo never > /sys/kernel/mm/transparent_hugepage/enabled

Disabling defrag also fixed the problem:

echo never > /sys/kernel/mm/transparent_hugepage/defrag

More discussion here:

http://structureddata.org/2012/06/18/linux-6-transparent-huge-pages-and-hadoop-workloads/ 

I suspect this will affect at least RHEL/CentOS 6 and recent Fedoras.

Any thoughts on how to make this better? It would actually be nice to have this big slice backed by hugepages on a THP-capable system.

Regards

Albert

[Kirill A. Shutemov]

On Wed, Jul 18, 2012 at 07:51:32AM -0700, Albert Strasheim wrote:
> Hello all
>
> We ran into an interesting problem today where a Go program making a few 1
> GB slices seems to spend forever allocating this memory (Linux, amd64).

Could you show code which demonstrate the issue?

--
Kirill A. Shutemov

[Albert Strasheim]

Hello

I was triggering this consistently today, however not with Go code, but with a Go build that was linking lots of commands:

http://code.google.com/p/go/issues/detail?id=3080#c13

Regards

Albert