On (23/04/17 19:53), Yosry Ahmed wrote: > > As for this patch, I personally do not observe a lot of compaction in > our production environment, and allowing one thread to perform > compaction while others move on with their lives can be better than > having all of them continuously contending for the pool->lock, which > means more contention with ~all zsmalloc operations, not just > concurrent compactors. I can't say for sure that this is an > improvement, but I *believe* it is. Looking at one of ChromeOS memory-pressure tests, I see that sometimes (albeit rarely) we can have up to 9 parallel zspool compaction contexts, perhaps a little bit too many for a 12 CPUs laptop: [ 2159.378827] zsmalloc: ctx #1 chrome -> zs_compact() [ 2159.379002] zsmalloc: ctx #2 Chrome_ChildIOT -> zs_compact() [ 2159.379120] zsmalloc: ctx #3 chrome -> zs_compact() [ 2159.379135] zsmalloc: ctx #4 chrome -> zs_compact() [ 2159.379213] zsmalloc: ctx #5 chrome -> zs_compact() [ 2159.379271] zsmalloc: ctx #6 chrome -> zs_compact() [ 2159.379276] zsmalloc: ctx #7 chrome -> zs_compact() [ 2159.382786] zsmalloc: ctx #8 chrome -> zs_compact() [ 2159.432153] zsmalloc: ctx #9 kswapd0 -> zs_compact()
