On Thu, Apr 30, 2015 at 01:11:39PM +0300, Alberto Garcia wrote: > Here are some improvements to the qcow2 L2/refcount cache code. > > The first one is that all cache tables are now allocated using a > single memory block, as we discussed last week. > > Apart from a more efficient use of memory, this allows some additional > optimizations so I took the chance to make other changes. > > - qcow2_cache_put() and qcow2_cache_entry_mark_dirty() are now O(1) > - The eviction algorithm is now LRU. The previous one only works well > with very small cache sizes. > - qcow2_cache_find_entry_to_replace() is no longer necessary. > - Lookups are faster now. > > In my tests with a preallocated 128MB L2 cache in an empty drive the > new code is ~13% faster than the previous one (~43% if compiled > without optimizations). This is a best-case scenario, if the cache is > smaller or the drive is full of data the improvements are not so > visible, but I believe the code is simpler now so I hope you find the > changes worthwhile.
Overall I'm happy with this approach. Looking forward to the next revision. Stefan
pgp6cAGQFwxbD.pgp
Description: PGP signature
