Messages in this thread | | | Date | Fri, 24 Jul 2015 10:12:37 -0700 | From | Jörn Engel <> | Subject | Re: [PATCH] hugetlb: cond_resched for set_max_huge_pages and follow_hugetlb_page |
| |
On Fri, Jul 24, 2015 at 08:59:59AM +0200, Michal Hocko wrote: > On Thu 23-07-15 14:54:31, Spencer Baugh wrote: > > From: Joern Engel <joern@logfs.org> > > > > ~150ms scheduler latency for both observed in the wild. > > This is way to vague. Could you describe your problem somehow more, > please? > There are schduling points in the page allocator (when it triggers the > reclaim), why are those not sufficient? Or do you manage to allocate > many hugetlb pages without performing the reclaim and that leads to > soft lockups?
We don't use transparent hugepages - they cause too much latency. Instead we reserve somewhere around 3/4 or so of physical memory for hugepages. "sysctl -w vm.nr_hugepages=100000" or something similar in a startup script.
Since it is early in boot we don't go through page reclaim.
Jörn
-- Everything should be made as simple as possible, but not simpler. -- Albert Einstein
| |