On Fri, Aug 08, 2014 at 02:32:58PM +0200, Michal Hocko wrote: > On Thu 07-08-14 11:31:41, Johannes Weiner wrote: > > On Thu, Aug 07, 2014 at 03:08:22PM +0200, Michal Hocko wrote: > > > On Mon 04-08-14 17:14:54, Johannes Weiner wrote: > > > > Instead of passing the request size to direct reclaim, memcg just > > > > manually loops around reclaiming SWAP_CLUSTER_MAX pages until the > > > > charge can succeed. That potentially wastes scan progress when huge > > > > page allocations require multiple invocations, which always have to > > > > restart from the default scan priority. > > > > > > > > Pass the request size as a reclaim target to direct reclaim and leave > > > > it to that code to reach the goal. > > > > > > THP charge then will ask for 512 pages to be (direct) reclaimed. That > > > is _a lot_ and I would expect long stalls to achieve this target. I > > > would also expect quick priority drop down and potential over-reclaim > > > for small and moderately sized memcgs (e.g. memcg with 1G worth of pages > > > would need to drop down below DEF_PRIORITY-2 to have a chance to scan > > > that many pages). All that done for a charge which can fallback to a > > > single page charge. > > > > > > The current code is quite hostile to THP when we are close to the limit > > > but solving this by introducing long stalls instead doesn't sound like a > > > proper approach to me. > > > > THP latencies are actually the same when comparing high limit nr_pages > > reclaim with the current hard limit SWAP_CLUSTER_MAX reclaim, > > Are you sure about this? I fail to see how they can be same as THP > allocations/charges are __GFP_NORETRY so there is only one reclaim > round for the hard limit reclaim followed by the charge failure if > it is not successful. I use this test program that faults in anon pages, reports average and max for every 512-page chunk (THP size), then reports the aggregate at the end: memory.max: avg=18729us max=450625us real 0m14.335s user 0m0.157s sys 0m6.307s memory.high: avg=18676us max=457499us real 0m14.375s user 0m0.046s sys 0m4.294s > > although system time is reduced with the high limit. > > High limit reclaim with SWAP_CLUSTER_MAX has better fault latency but > > it doesn't actually contain the workload - with 1G high and a 4G load, > > the consumption at the end of the run is 3.7G. > > Wouldn't it help to simply fail the charge and allow the charger to > fallback for THP allocations if the usage is above high limit too > much? The follow up single page charge fallback would be still > throttled. This is about defining the limit semantics in unified hierarchy, and not really the time or place to optimize THP charge latency. What are you trying to accomplish here? > > So what I'm proposing works and is of equal quality from a THP POV. > > This change is complicated enough when we stick to the facts, let's > > not make up things based on gut feeling. > > Agreed and I would expect those _facts_ to be part of the changelog. You made unfounded claims about THP allocation latencies, and I showed the numbers to refute it, but that doesn't make any of this relevant for the changelogs of these patches. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>