Hi, On 2020-11-17 00:09:34 +0300, Marina Polyakova wrote: > Sorry I'm not familiar with the internal architecture of snapshots, locks > etc. in postgres, but I wanted to ask - what exact kind of patch for fair > lwlocks do you want to offer to the community? I applied the 6-th version of > the patch for fair lwlocks from [1] to the old master branch (see commit > [2]), started many threads in pgbench (-M prepared -c 1000 -j 500 -T 10 -P1 > -S) and I did not receive stable first progress reports, which IIUC are one > of the advantages of the discussed patch for the pgbench (see [3])...
Thanks for running some benchmarks. I think it's quite unsurprising that you'd see skewed numbers initially, even with fairer locks. Just by virtue of pgbench starting threads and each thread immediately starting to perform work, you are bound to get odd pretty meaningless initial numbers. Even without contention, and when using fewer connections than there are CPUs. And especially when starting a larger number of connections, because the main pgbench thread will get fewer and fewer scheduler slices because of the threads running benchmarks already. Regards, Andres
