On Fri, 16 Sept 2022 at 17:08, Michail Nikolaev <michail.nikol...@gmail.com> wrote: > > Hello everyone. > > To find the best frequency for calling KnownAssignedXidsCompress in > Simon's patch, I made a set of benchmarks. It looks like each 8th xid > is a little bit often. > > Setup and method is the same as previous (1). 16-core machines, > max_connections = 5000. Tests were running for about a day, 220 runs > in total (each version for 20 times, evenly distributed throughout the > day). > > Staring from 60th second, 30 seconds-long transaction was started on primary. > > Graphs in attachment. So, looks like 64 is the best value here. It > gives even a little bit more TPS than smaller values. > > Yes, such benchmark does not cover all possible cases, but it is > better to measure at least something when selecting constants :)
This is very good and clear, thank you. > If someone has an idea of different benchmark scenarios - please share them. > So, updated version (with 64 and some commit message) in attachment too. > > [1]: > https://www.postgresql.org/message-id/flat/CANtu0ohzBFTYwdLtcanWo4%2B794WWUi7LY2rnbHyorJdE8_ZnGg%40mail.gmail.com#379c1be7b8134ada5a574078d51b64c6 I've cleaned up the comments and used a #define for the constant. IMHO this is ready for commit. I will add it to the next CF. -- Simon Riggs http://www.EnterpriseDB.com/
v8c-new-heuristic-to-compress-KnownAssignedXids.patch
Description: Binary data
