On the cache-server branch: Eliminate 70+% of the cache access overhead
caused by calculating expensive hash sums.
We use two approaches here. First, cache the combined_key (i.e. hash)
values from the last access and skip the calculation if the same key is
being used again. This is very effective because we often probe a cache
with a specific key and then set a value for that key.
Secondly, we leverage the fact that most keys in fsfs-format7 code will
be fixed-size 16 byte keys (80+% up from ~50%). Simply combining those
with the respective cache prefix will already produce a unique key. We
add some scrambling code to distribute their values somewhat more evenly
over the key space.
* subversion/libsvn_subr/cache-membuffer.c
(): update / correct global doc
(get_group_index): compensate for the worsened key value distribution
(last_access_key_t): new data structure
(svn_membuffer_cache_t): add cache for the last combine_key() result
(combine_long_key): split from combine_key(); use cached values
(combine_key): split; add fast code for short fixed-size keys
(svn_cache__create_membuffer_cache): update
git-svn-id: https://svn.apache.org/repos/asf/subversion/branches/cache-server@1458643 13f79535-47bb-0310-9956-ffa450edef68
1 file changed
