Hi, On 2021-02-18 16:09:49 +1300, Thomas Munro wrote: > In another thread[1], I proposed $SUBJECT, but then we found a better > solution to that thread's specific problem. The general idea is still > good though: it's possible to (1) replace several existing copies of > our qsort algorithm with one, and (2) make new specialised versions a > bit more easily than the existing Perl generator allows. So, I'm back > with a rebased stack of patches. I'll leave specific cases for new > worthwhile specialisations for separate proposals; I've heard about > several.
One place that could benefit is the qsort that BufferSync() does at the start. I tried your patch for that, and it does reduce the sort time considerably. For 64GB of mostly dirty shared_buffers from ~1.4s to 0.6s. Now, obviously one can argue that that's not going to be the crucial spot, and wouldn't be entirely wrong. OTOH, in my AIO branch I see checkpointer doing ~10GB/s, leading to the sort being a measurable portion of the overall time. Greetings, Andres Freund
