On Tue, Mar 15, 2016 at 10:33:03AM +0900, Sergey Senozhatsky wrote: > On (03/15/16 09:46), Minchan Kim wrote: > [..] > > > yes, > > > > > > we do less work this way - scan and compact only one class, instead > > > of locking and compacting all of them; which sounds reasonable. > > > > Hmm,, It consumes more memory(i.e., sizeof(work_struct) + sizeof(void *) > > + sizeof(bool) * NR_CLASS) as well as kicking many work up to NR_CLASS. > > yes, it does. not really happy with it either. > > > I didn't test your patch but I guess I can make worst case scenario. > > > > * make every class fragmented under 40% > > * On the 40% boundary, repeated alloc/free of every class so every free > > can schedule work if it was not scheduled. > > * Although class fragment is too high, it's not a problem if the class > > consumes small amount of memory. > > hm, in this scenario both solutions are less than perfect. we jump > X times over 40% margin, we have X*NR_CLASS compaction scans in the > end. the difference is that we queue less works, yes, but we don't > have to use workqueue in the first place; compaction can be done > asynchronously by a pool's dedicated kthread. so we will just > wake_up() the process. Hmm, kthread is over-engineered to me. If we want to create new kthread in the system, I guess we should persuade many people to merge in. Surely, we should have why it couldn't be done by others(e.g., workqueue). I think your workqueue approach is good to me. Only problem I can see with it is we cannot start compaction when we want instantly so my conclusion is we need both direct and background compaction. For shrinker and user-space trigger knob, we could compact in that context while we could queue background job to compact in zs_free. > > > I guess it can make degradation if I try to test on zsmalloc > > microbenchmark. > > > > As well, although I don't know workqueue internal well, thesedays, > > I saw a few of mails related to workqueue(maybe, vmstat) and it had > > some trouble if system memory pressure is heavy IIRC. > > yes, you are right. wq provides WQ_MEM_RECLAIM bit for this > case -- a special kthread that it will wake up to process works. > > > My approach is as follows, for exmaple. > > > > Let's make a global ratio. Let's say it's 4M. > > ok. should it depend on pool size? min(20% of pool_size, XXMB)? Maybe, that could be a knob but need to think more what should be default. In this moment, clear thing is that we should prevent frequent ping-pong background compaction as repeated alloc/free with dancing on threshold boundary. > > > If zs_free(or something) realizes current fragment is over 4M, > > kick compacion backgroud job. > > yes, zs_free() is the only place that introduces fragmentation. > > > The job scans from highest to lower class and compact zspages > > in each size_class until it meets high watermark(e.g, 4M + 4M /2 = > > 6M fragment ratio). > > ok. > > > And in the middle of background compaction, if we find it's too > > many scan(e.g., 256 zspages or somethings), just bail out the > > job for the latency and reschedule it for next time. At the next > > time, we can continue from the last size class. > > ok. I'd probably prefer more simple rules here: > -- bail out because it has compacted XXMB > so the fragmentation ratio is *expected* to be below the watermark Need high watermark to stop compaction. It will prevent frequent background compaction triggering. > -- nothing to scan anymore > compaction is executed concurrently with zs_free()/zs_malloc() > calls, it's harder to control/guarantee some global state. > > overall, no real objections. this approach can work, I think. need > to test it. Thanks, Sergey! -- 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>