Excerpts from Linus Torvalds's message of March 18, 2021 5:26 am: > On Wed, Mar 17, 2021 at 3:44 AM Nicholas Piggin <npiggin@xxxxxxxxx> wrote: >> >> Argh, because I didn't test small. Sorry I had the BASE_SMALL setting in >> another patch and thought it would be a good idea to mash them together. >> In hindsight probably not even if it did build. > > I was going to complain about that code in general. > > First complaining about the hash being small, and then adding a config > option to make it ridiculously much *smaller* seemed wrong to begin > with, and didn't make any sense. > > So no, please don't smash together. Fair point, fixed. > > In fact, I'd like to see this split up, and with more numbers: > > - separate out the bit_waitqueue thing that is almost certainly not > remotely as critical (and maybe not needed at all) > > - show the profile number _after_ the patch(es) Might take some time to get a system and run tests. We actually had difficulty recreating it before this patch too, so it's kind of hard to say _that_ was the exact case that previously ran badly and is now fixed. We thought just the statistical nature of collisions and page / lock contention made things occasionally line up and tank. > - explain why you picked the random scaling numbers (21 and 22 for > the two different cases)? > > - give an estimate of how big the array now ends up being for > different configurations. > > I think it ends up using that "scale" factor of 21, and basically > being "memory size >> 21" and then rounding up to a power of two. > > And honestly, I'm not sure that makes much sense. So for a 1GB machine > we get the same as we used to for the bit waitqueue (twice as many for > the page waitqueue) , but if you run on some smaller setup, you > apparently can end up with just a couple of buckets. > > So I'd feel a lot better about this if I saw the numbers, and got the > feeling that the patch actually tries to take legacy machines into > account. > > And even on a big machine, what's the advantage of scaling perfectly > with memory. If you have a terabyte of RAM, why would you need half a > million hash entries (if I did the math right), and use 4GB of memory > on it? The contention doesn't go up by amount of memory, it goes up > roughly by number of threads, and the two are very seldom really all > that linearly connected. > > So honestly, I'd like to see more reasonable numbers. I'd like to see > what the impact of just raising the hash bit size from 8 to 16 is on > that big machine. Maybe still using alloc_large_system_hash(), but > using a low-imit of 8 (our traditional very old number that hasn't > been a problem even on small machines), and a high-limit of 16 or > something. > > And if you want even more, I really really want that justified by the > performance / profile numbers. Yes all good points I'll add those numbers. It may need a floor and ceiling or something like that. We may not need quite so many entries. > > And does does that "bit_waitqueue" really merit updating AT ALL? It's > almost entirely unused these days. I updated it mainly because keeping the code more similar ends up being easier than unnecessary diverging. The memory cost is no big deal (once limits are fixed) so I prefer not to encounter some case where it falls over. > I think maybe the page lock code > used to use that, but then realized it had more specialized needs, so > now it's separate. > > So can we split that bit-waitqueue thing up from the page waitqueue > changes? They have basically nothing in common except for a history, > and I think they should be treated separately (including the > explanation for what actually hits the bottleneck). It's still used. Buffer heads being an obvious and widely used one that follows similar usage pattern as page lock / writeback in some cases. Several other filesystems seem to use it for similar block / IO tracking structures by the looks (md, btrfs, nfs). Thanks, Nick
