On Fri, Jan 18, 2019 at 06:07:45PM +0100, Paolo Valente wrote: > > > > Il giorno 18 gen 2019, alle ore 17:35, Josef Bacik <josef@xxxxxxxxxxxxxx> ha scritto: > > > > On Fri, Jan 18, 2019 at 11:31:24AM +0100, Andrea Righi wrote: > >> This is a redesign of my old cgroup-io-throttle controller: > >> https://lwn.net/Articles/330531/ > >> > >> I'm resuming this old patch to point out a problem that I think is still > >> not solved completely. > >> > >> = Problem = > >> > >> The io.max controller works really well at limiting synchronous I/O > >> (READs), but a lot of I/O requests are initiated outside the context of > >> the process that is ultimately responsible for its creation (e.g., > >> WRITEs). > >> > >> Throttling at the block layer in some cases is too late and we may end > >> up slowing down processes that are not responsible for the I/O that > >> is being processed at that level. > > > > How so? The writeback threads are per-cgroup and have the cgroup stuff set > > properly. So if you dirty a bunch of pages, they are associated with your > > cgroup, and then writeback happens and it's done in the writeback thread > > associated with your cgroup and then that is throttled. Then you are throttled > > at balance_dirty_pages() because the writeout is taking longer. > > > > IIUC, Andrea described this problem: certain processes in a certain group dirty a > lot of pages, causing write back to start. Then some other blameless > process in the same group experiences very high latency, in spite of > the fact that it has to do little I/O. > > Does your blk_cgroup_congested() stuff solves this issue? > > Or simply I didn't get what Andrea meant at all :) > > Thanks, > Paolo Yes, there is also this problem: provide fairness among processes running inside the same cgroup. This is a tough one, because once the I/O limit is reached whoever process comes next gets punished, even if it hasn't done any I/O before. Well, my proposal doesn't solve this problem. To solve this one in the "throttling" scenario, we should probably save some information about the previously generated I/O activity and apply a delay proportional to that (like a dynamic weight for each process inside each cgroup). AFAICS the io.max controller doesn't solve this problem either. -Andrea
