Hi, Pavel! On Fri, Nov 11, 2022 at 2:40 PM Pavel Borisov <pashkin.e...@gmail.com> wrote: > I've done some more measurements to check the hypotheses regarding the > performance of a previous patch v2, and explain the results of tests > in [1]. > > The results below are the same (tps vs connections) plots as in [1], > and the test is identical to the insert test in this thread [2]. > Additionally, in each case, there is a plot with results relative to > Andres Freund's patch [3]. Log plots are good for seeing details in > the range of 20-30 connections, but they somewhat hide the fact that > the effect in the range of 500+ connections is much more significant > overall, so I'd recommend looking at the linear plots as well.
Thank you for doing all the experiments! BTW, sometimes it's hard to distinguish so many lines on a jpg picture. Could I ask you to post the same graphs in png and also post raw data in csv format? > I'm also planning to do the same tests on an ARM server when the free > one comes available to me. > Thoughts? ARM tests should be great. We definitely need to check this on more than just one architecture. Please, check with and without LSE instructions. They could lead to dramatic speedup [1]. Although, most of precompiled binaries are distributed without them. So, both cases seems important to me so far. >From what we have so far, I think we could try combine the multiple strategies to achieve the best result. 2x1ms is one of the leaders before ~200 connections, and 1x1ms is once of the leaders after. We could implement simple heuristics to switch between 1 and 2 retries similar to what we have to spin delays. But let's have ARM results first. Links 1. https://akorotkov.github.io/blog/2021/04/30/arm/ ------ Regards, Alexander Korotkov