On Wed, Oct 13, 2010 at 12:25:26AM +0800, Mel Gorman wrote: > > > > > > In a 4 socket 64 CPU system, zone_nr_free_pages() takes about 5% ~ 10% cpu time > > > > > > according to perf when memory pressure is high. The workload does something > > > > > > like: > > > > > > for i in `seq 1 $nr_cpu` > > > > > > do > > > > > > create_sparse_file $SPARSE_FILE-$i $((10 * mem / nr_cpu)) > > > > > > $USEMEM -f $SPARSE_FILE-$i -j 4096 --readonly $((10 * mem / nr_cpu)) & > > > > > > done > > > > > > this simply reads a sparse file for each CPU. Apparently the > > > > > > zone->percpu_drift_mark is too big, and guess zone_page_state_snapshot() makes > > > > > > a lot of cache bounce for ->vm_stat_diff[]. below is the zoneinfo for reference. > > > > > > > > > > Would it be possible for you to post the oprofile report? I'm in the > > > > > early stages of trying to reproduce this locally based on your test > > > > > description. The first machine I tried showed that zone_nr_page_state > > > > > was consuming 0.26% of profile time with the vast bulk occupied by > > > > > do_mpage_readahead. See as follows > > > > > > > > > > 1599339 53.3463 vmlinux-2.6.36-rc7-pcpudrift do_mpage_readpage > > > > > 131713 4.3933 vmlinux-2.6.36-rc7-pcpudrift __isolate_lru_page > > > > > 103958 3.4675 vmlinux-2.6.36-rc7-pcpudrift free_pcppages_bulk > > > > > 85024 2.8360 vmlinux-2.6.36-rc7-pcpudrift __rmqueue > > > > > 78697 2.6250 vmlinux-2.6.36-rc7-pcpudrift native_flush_tlb_others > > > > > 75678 2.5243 vmlinux-2.6.36-rc7-pcpudrift unlock_page > > > > > 68741 2.2929 vmlinux-2.6.36-rc7-pcpudrift get_page_from_freelist > > > > > 56043 1.8693 vmlinux-2.6.36-rc7-pcpudrift __alloc_pages_nodemask > > > > > 55863 1.8633 vmlinux-2.6.36-rc7-pcpudrift ____pagevec_lru_add > > > > > 46044 1.5358 vmlinux-2.6.36-rc7-pcpudrift radix_tree_delete > > > > > 44543 1.4857 vmlinux-2.6.36-rc7-pcpudrift shrink_page_list > > > > > 33636 1.1219 vmlinux-2.6.36-rc7-pcpudrift zone_watermark_ok > > > > > ..... > > > > > 7855 0.2620 vmlinux-2.6.36-rc7-pcpudrift zone_nr_free_pages > > > > > > > > > > The machine I am testing on is non-NUMA 4-core single socket and totally > > > > > different characteristics but I want to be sure I'm going more or less the > > > > > right direction with the reproduction case before trying to find a larger > > > > > machine. > > > > > > > > Here it is. this is a 4 socket nahalem machine. > > > > 268160.00 57.2% _raw_spin_lock /lib/modules/2.6.36-rc5-shli+/build/vmlinux > > > > 40302.00 8.6% zone_nr_free_pages /lib/modules/2.6.36-rc5-shli+/build/vmlinux > > > > 36827.00 7.9% do_mpage_readpage /lib/modules/2.6.36-rc5-shli+/build/vmlinux > > > > 28011.00 6.0% _raw_spin_lock_irq /lib/modules/2.6.36-rc5-shli+/build/vmlinux > > > > 22973.00 4.9% flush_tlb_others_ipi /lib/modules/2.6.36-rc5-shli+/build/vmlinux > > > > 10713.00 2.3% smp_invalidate_interrupt /lib/modules/2.6.36-rc5-shli+/build/vmlinux > > > > > > <SNIP> > > > > > Basically the similar test. I'm using Fengguang's test, please check attached > > file. I didn't enable lock stat or debug. The difference is my test is under a > > 4 socket system. In a 1 socket system, I don't see the issue too. > > > > Ok, finding a large enough machine was key here true enough. I don't > have access to Nehalem boxes but the same problem showed up on a large > ppc64 machine (8 socket, interestingly enough a 3 socket did not have any > significant problem). Based on that, I reproduced the problem and came up > with the patch below. > > Christoph, can you look at this please? I know you had concerns about adjusting > thresholds as being an expensive operation but the patch limits how often it > occurs and it seems better than reducing thresholds for the full lifetime of > the system just to avoid counter drift. What I did find with the patch that > the overhead of __mod_zone_page_state() increases because of the temporarily > reduced threshold. It goes from 0.0403% of profile time to 0.0967% on one > machine and from 0.0677% to 0.43% on another. As this is just while kswapd > is awake, it seems withiin an acceptable margin but it is a caution against > simply reducing the existing thresholds. What is more relevant is the time > to complete the benchmark is increased due to the reduction of the thresholds. > This is a tradeoff between being fast and safe but I'm open to > suggestions on how high a safe threshold might be. > > Shaohua, can you test keeping an eye out for any additional function > that is now taking a lot more CPU time? seems ok so far in the 4 sockets system. In this system, each node has 8G memory, so the threshold is 5 with memory pressure. Haven't tested this in some small machines yet, for example, each node just has 4G memory, etc. Thanks, Shaohua -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>