On Wed, Nov 23, 2022 at 12:11:24AM -0800, Yosry Ahmed wrote: > On Wed, Nov 23, 2022 at 12:02 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote: > > On Tue, Nov 22, 2022 at 7:50 PM Sergey Senozhatsky > > > There are no accesses to swapped out pages yes, but zspage holds multiple > > > objects, which are compressed swapped out pages in this particular case. > > > For example, zspage in class size 176 (bytes) can hold 93 objects per-zspage, > > > that is 93 compressed swapped out pages. Consider ZS_FULL zspages which > > > is at the tail of the LRU list. Suppose that we page-faulted 20 times and > > > read 20 objects from that zspage, IOW zspage has been in use 20 times very > > > recently, while writeback still considers it to be "not-used" and will > > > evict it. > > > > > > So if this works for you then I'm fine. But we probably, like you suggested, > > > can document a couple of things here - namely why WRITE access to zspage > > > counts as "zspage is in use" but READ access to the same zspage does not > > > count as "zspage is in use". > Nonetheless, the fact that we refaulted an object in a zspage does not > necessarily mean that other objects on the same are hotter than > objects in other zspages IIUC. Yes. On allocation, we know that there is at least one hot object in the page. On refault, the connection between objects in a page is weak. And it's weaker on zsmalloc than with other backends due to the many size classes making temporal grouping less likely. So I think you're quite right, Segey, that a per-class LRU would be more accurate. It's no-LRU < zspage-LRU < class-LRU < object-LRU. Like Yosry said, the plan is to implement an object-LRU next as part of the generalized LRU for zsmalloc, zbud and z3fold. For now, the zspage LRU is an improvement to no-LRU. Our production experiments confirmed that.
