Re: [bug?] mallocstress poor performance with THP on arm64 system

Hi Kirill,

"Kirill A. Shutemov" [off-list ref] writes:

On Wed, Feb 14, 2018 at 06:05:01PM -0500, Jan Stancek wrote:

quoted

Hi,

mallocstress[1] LTP testcase takes ~5+ minutes to complete
on some arm64 systems (e.g. 4 node, 64 CPU, 256GB RAM):
 real    7m58.089s
 user    0m0.513s
 sys     24m27.041s

But if I turn off THP ("transparent_hugepage=never") it's a lot faster:
 real    0m4.185s
 user    0m0.298s
 sys     0m13.954s

It's multi-threaded workload. My *guess* is that poor performance is due
to lack of ARCH_ENABLE_SPLIT_PMD_PTLOCK support on arm64.

In this instance I think the latency is due to the large size of PMD
hugepages and THP=always. But split PMD locks seem like a useful feature
to have for large core count systems.

I'll have a go at enabling this for arm64.

Thanks,
Punit

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>