ben-manes / caffeine / #5781

100%

master: 100%

push

github

avoid false sharing within the fastpath optimization check

Every hit on a size-bounded cache with strong references and no access
expiration tested whether the frequency sketch was initialized, which
loads sketch.table. That field shares a cache line with the sketch's
size, which the maintenance thread increments for nearly every drained
read, so the reader kept reloading a line that had just been written.

Such a cache now starts with the disabled read buffer and swaps in a
BoundedBuffer wherever the sketch is initialized: at half occupancy,
when presized, or when setMaximum makes it half full. The hit path
offers to the buffer without testing the sketch. On a read-only sweep
of a presized 65,536-entry cache this took a hit from 11.65 to 9.99 ns
at one thread (~16% speedup) and from 11.13 to 7.00 ns per thread at
eight (~60% speedup), and a cache under half full was no slower.

4714 of 4778 branches covered (98.66%)

8 of 8 new or added lines in 1 file covered. (100.0%)

9167 of 9179 relevant lines covered (99.87%)

1.0 hits per line