On Wed, Apr 22, 2015 at 04:37:15PM +0200, Alberto Garcia wrote: > On Wed 22 Apr 2015 12:26:02 PM CEST, Stefan Hajnoczi wrote: > > >> Large disk images need large L2 caches in order to maximize their I/O > >> performance. However setting a correct size for the cache is not > >> necessarily easy since apart from the image size, it also depends on > >> other factors like its usage patterns or whether it's part of a > >> backing chain. > >> > >> In order to be able to set a very large cache size to cover the > >> worst-case scenarios and yet prevent an unnecessary waste of memory, > >> this patch modifies the qcow2 cache algorithm so the memory for each > >> entry is allocated only when it's actually needed. > >> > >> This also improves the scenarios with smaller images: the current > >> default L2 cache size can map a whole 8GB disk, so those smaller than > >> that are allocating cache memory that can never be used. > > > > What measurable improvement does this patch make? > > > > Buffers allocated upfront are not touched, so they don't actually > > consume physical memory. > > For a cache size of 128MB, the PSS is actually ~10MB larger without the > patch, which seems to come from posix_memalign().
Do you mean RSS or are you using a tool that reports a "PSS" number that I don't know about? We should understand what is going on instead of moving the code around to hide/delay the problem. Stefan
pgp1X8ELVEjkj.pgp
Description: PGP signature
