timescale/promscale

Consider removing [16]byte padding from cache element

niksajakovljevic opened this issue · 0 comments

Currently each cache element is padded with [16]byte. This has been done to prevent CPU cache lines bouncing (happening during concurrent access) by making sure that element occupies 64 bytes (a typical cache line size)(check #1705 for detailed discussion).
However some recent benchmarks are showing almost no difference (or very little difference in single digit %) for padded code version. Basically the new findings are pointing in direction of removing [16]byte padding. This will greatly reduce memory usage since we are caching millions of elements.