On Wed, 2018-07-11 at 15:07 -0400, Waiman Long wrote: > On 07/11/2018 01:42 PM, James Bottomley wrote: > > On Wed, 2018-07-11 at 11:13 -0400, Waiman Long wrote: > > > On 07/11/2018 06:21 AM, Michal Hocko wrote: > > > > On Tue 10-07-18 12:09:17, Waiman Long wrote: > > > > [...] > > > > > I am going to reduce the granularity of each unit to 1/1000 > > > > > of the total system memory so that for large system with TB > > > > > of memory, a smaller amount of memory can be specified. > > > > > > > > It is just a matter of time for this to be too coarse as well. > > > > > > The goal is to not have too much memory being consumed by > > > negative > > > dentries and also the limit won't be reached by regular daily > > > activities. So a limit of 1/1000 of the total system memory will > > > be good enough on large memory system even if the absolute number > > > is really big. > > > > OK, I think the reason we're going round and round here without > > converging is that one of the goals of the mm subsystem is to > > manage all of our cached objects and to it the negative (and > > positive) dentries simply look like a clean cache of > > objects. Right at the moment mm manages them in the same way it > > manages all the other caches, a lot of which suffer from the "you > > can cause lots of allocations to artificially grow them" > > problem. So the main question is why doesn't the current mm > > control of the caches work well enough for dentries? > > What are the problems you're seeing that mm should be catching? If > > you can answer this, then we could get on to whether a separate > > shrinker, cache separation or some fix in mm itself is the right > > answer. > > > > What you say above is based on a conclusion: limiting dentries > > improves the system performance. What we're asking for is evidence > > for that conclusion so we can explore whether the same would go for > > any of our other system caches (so do we have a global cache > > management problem or is it only the dentry cache?) > > > > James > > > > I am not saying that limiting dentries will improve performance. I am > just saying that unlimited growth in the number of negative dentries > will reduce the amount of memory available to other applications and > hence will have an impact on performance. Normally the amount of > memory consumed by dentries is a very small portion of the system > memory. OK, can we poke on only this point for a while? Linux never really has any "available memory": pretty much as soon as you boot up the system will consume all your free memory for some type of cache (usually the page cache which got filled during boot). The expectation is that in a steady, running, state the system is using almost all available memory for caching something ... if it's not negative dentries it will be something else. The mm assumption is that clean cache is so cheap to recover that it's almost equivalent to free memory and your patch is saying this isn't so and we have a problem dumping the dentry cache. So, why can't we treat the dentry cache as equivalent to free memory? What in your tests is making it harder to recover the memory in the dentry cache? James
