Hi, On 2018-06-24 22:32:07 +0200, Dmitry Dolgov wrote: > `perf diff` indeed shows that in the first case (with the 4M rows dataset) the > jitted version has some noticeable delta for one call, and unfortunately so > far > I couldn't figure out which one exactly because of JIT (btw, who can explain > how to see a correct full `perf report` in this case? Somehow `perf > inject --jit -o > perf.data.jitted` and jit_profiling_support didn't help).
jit_profiling_support currently requires a patch (that I'm about to merge into their trunk) in LLVM. I've posted it a couple times to the list. > But since on the bigger dataset I've got expected results, maybe it's just a > sign that JIT kicks in too early in this case and what's necessary is to > adjust > jit_above_cost/jit_optimize_above_cost/jit_inline_above_cost? Yea, that's very likely the problem. I wonder if we should multiply the cost w/ the number of targeted workers to reduce problems with parallelism... Greetings, Andres Freund