On 2014/6/23 19:29, Michal Hocko wrote: > On Mon 23-06-14 10:05:48, Xishi Qiu wrote: >> On 2014/6/20 23:32, Michal Hocko wrote: >> >>> On Fri 20-06-14 15:56:56, Xishi Qiu wrote: >>>> On 2014/6/17 9:35, Xishi Qiu wrote: >>>> >>>>> On 2014/6/16 20:50, Rafael Aquini wrote: >>>>> >>>>>> On Mon, Jun 16, 2014 at 01:14:22PM +0200, Michal Hocko wrote: >>>>>>> On Mon 16-06-14 17:24:38, Xishi Qiu wrote: >>>>>>>> When system(e.g. smart phone) running for a long time, the cache often takes >>>>>>>> a large memory, maybe the free memory is less than 50M, then OOM will happen >>>>>>>> if APP allocate a large order pages suddenly and memory reclaim too slowly. >>>>>>> >>>>>>> Have you ever seen this to happen? Page cache should be easy to reclaim and >>>>>>> if there is too mach dirty memory then you should be able to tune the >>>>>>> amount by dirty_bytes/ratio knob. If the page allocator falls back to >>>>>>> OOM and there is a lot of page cache then I would call it a bug. I do >>>>>>> not think that limiting the amount of the page cache globally makes >>>>>>> sense. There are Unix systems which offer this feature but I think it is >>>>>>> a bad interface which only papers over the reclaim inefficiency or lack >>>>>>> of other isolations between loads. >>>>>>> >>>>>> +1 >>>>>> >>>>>> It would be good if you could show some numbers that serve as evidence >>>>>> of your theory on "excessive" pagecache acting as a trigger to your >>>>>> observed OOMs. I'm assuming, by your 'e.g', you're running a swapless >>>>>> system, so I would think your system OOMs are due to inability to >>>>>> reclaim anon memory, instead of pagecache. >>>>>> >>>> >>>> I asked some colleagues, when the cache takes a large memory, it will not >>>> trigger OOM, but performance regression. >>>> >>>> It is because that business process do IO high frequency, and this will >>>> increase page cache. When there is not enough memory, page cache will >>>> be reclaimed first, then alloc a new page, and add it to page cache. This >>>> often takes too much time, and causes performance regression. >>> >>> I cannot say I would understand the problem you are describing. So the >>> page cache eats the most of the memory and that increases allocation >>> latency for new page cache? Is it because of the direct reclaim? >> >> Yes, allocation latency causes performance regression. > > This doesn't make much sense to me. So you have a problem with latency > caused by direct reclaim so you add a new way of direct page cache > reclaim. > >> A user process produces page cache frequently, so free memory is not >> enough after running a long time. Slow path takes much more time because >> direct reclaim. And kswapd will reclaim memory too, but not much. Thus it >> always triggers slow path. this will cause performance regression. > > If I were you I would focus on why the reclaim doesn't catch up with the > page cache users. The mechanism you are proposing in unacceptable. Hi Michal, Do you mean why the reclaim is slower than page cache increase? I think there are two reasons: 1. kswapd and direct_reclaim will be triggered only when there is not enough memory(e.g. __alloc_pages_slowpath()). That means it will not reclaim when memory is enough(e.g. get_page_from_freelist()). 2. __alloc_pages_direct_reclaim try_to_free_pages nr_to_reclaim = SWAP_CLUSTER_MAX And "#define SWAP_CLUSTER_MAX 32UL", that means it expect to reclaim 32 pages. It is too few, if we alloc 2^10 pages in one time. Thanks, Xishi Qiu -- 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>