04:06 <phlebas> antocuni: i get the same relation, but note that we drop the first half of the iterations. warmup is very very slow for us

04:07 <phlebas> as in, my first couple of iterations are 200x slower. if those are included in the average, it's skewed

04:24 <phlebas> ah, and another thing is that we are running the enterprise edition graalvm compiler in our ci, but the public nightly releases are the community edition. there's a bit of peak performance difference there too (about another 1x on my local machine)

04:25 <phlebas> i can send you am enterprise nightly, these are free to use anyway, just not commercially

08:53 <antocuni> phlebas: I still get the same result (~1900 ms per iteration) even if I run the benchmark function twice in a row, so it seems that the warmup is not the culprit. I wonder whether I am doing anything wrong or there is a performance problem that you didn't expect

08:55 <antocuni> I'm using this patch on the hpy-0.0.2 branch of the ultrajson-hpy repo: https://paste.opendev.org/show/807545/

08:55 <antocuni> (note that it calls benchmark(usjon_hpy_universal,...) twice)

08:55 <antocuni> and get this: https://paste.opendev.org/show/807546/