On 3/16/20 4:46 PM, Shakeel Butt wrote:
On Mon, Mar 16, 2020 at 3:24 PM Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote:When backporting commit 9c4e6b1a7027 ("mm, mlock, vmscan: no more skipping pagevecs") to our 4.9 kernel, our test bench noticed around 10% down with a couple of vm-scalability's test cases (lru-file-readonce, lru-file-readtwice and lru-file-mmap-read). I didn't see that much down on my VM (32c-64g-2nodes). It might be caused by the test configuration, which is 32c-256g with NUMA disabled and the tests were run in root memcg, so the tests actually stress only one inactive and active lru. It sounds not very usual in mordern production environment. That commit did two major changes: 1. Call page_evictable() 2. Use smp_mb to force the PG_lru set visible It looks they contribute the most overhead. The page_evictable() is a function which does function prologue and epilogue, and that was used by page reclaim path only. However, lru add is a very hot path, so it sounds better to make it inline. However, it calls page_mapping() which is not inlined either, but the disassemble shows it doesn't do push and pop operations and it sounds not very straightforward to inline it. Other than this, it sounds smp_mb() is not necessary for x86 since SetPageLRU is atomic which enforces memory barrier already, replace it with smp_mb__after_atomic() in the following patch. With the two fixes applied, the tests can get back around 5% on that test bench and get back normal on my VM. Since the test bench configuration is not that usual and I also saw around 6% up on the latest upstream, so it sounds good enough IMHO. The below is test data (lru-file-readtwice throughput) against the v5.6-rc4: mainline w/ inline fix 150MB 154MB With this patch the throughput gets 2.67% up. The data with using smp_mb__after_atomic() is showed in the following patch. Fixes: 9c4e6b1a7027 ("mm, mlock, vmscan: no more skipping pagevecs") Cc: Shakeel Butt <shakeelb@xxxxxxxxxx> Cc: Vlastimil Babka <vbabka@xxxxxxx> Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx> Signed-off-by: Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx>So, I tested on a real machine with limiting the 'dd' on a single node and reading 100 GiB sparse file (less than a single node). I just ran a single instance to not cause the lru lock contention. The cmd I used is "dd if=file-100GiB of=/dev/null bs=4k". I ran the cmd 10 times with drop_caches in between and measured the time it took. Without patch: 56.64143 +- 0.672 sec With patches: 56.10 +- 0.21 sec Reviewed-and-Tested-by: Shakeel Butt <shakeelb@xxxxxxxxxx>
Thanks Shakeel. It'd better to add your test result in the commit log too.
