#asahi · 2022-02-08 18:08
axboehuh#asahi · 2022-02-08 18:08
axboeinterestingly, I get 2.48s reliably on the perf core, 2.22s on the E#asahi · 2022-02-08 18:05
axboesounds good#asahi · 2022-02-08 18:04
axboenull: smc/work branch, to be exact#asahi · 2022-02-08 18:04
axboenull: actually just using the asahi linux repo for now, but was planning on merging it with current -git#asahi · 2022-02-08 17:57
axboeMTE is enabled in the config#asahi · 2022-02-08 17:57
axboenull: 820msec vs 2.48 sec#asahi · 2022-02-08 17:55
axboemps: just make should do it#asahi · 2022-02-08 17:52
axboenull: yeah something like that, just running getpid and timing#asahi · 2022-02-08 17:52
axboebut should be trivial to write a bench for that, if need be :)#asahi · 2022-02-08 17:52
axboenot sure if lmtest is still a thing, it used to do syscall timing#asahi · 2022-02-08 17:51
axboeshould build it on linux#asahi · 2022-02-08 17:51
axboeare you on osx?#asahi · 2022-02-08 17:50
axboeno pti, retpoline, etc#asahi · 2022-02-08 17:50
axboeall the x86 security stuff is turned off, fwiw#asahi · 2022-02-08 17:49
axboevs batched submit+complete#asahi · 2022-02-08 17:49
axboe-N1 asks it to do NOPs, which are just a roundtrip basically. -c1 -s1 is "complete 1 at the time, submit 1 at the time"#asahi · 2022-02-08 17:48
axboemps: it's in the fio repo#asahi · 2022-02-08 17:48
axboemps: see above, t/io_uring -N1 -c1 -s1#asahi · 2022-02-08 17:48
axboe~65% time just doing the syscall#asahi · 2022-02-08 17:47
axboehttps://kernel.dk/m1-sys.jpg vs https://kernel.dk/x86-sys.jpg - gives a good idea of the difference#asahi · 2022-02-08 17:43
axboedisabling PAC is good for about a 7.5% improvement#asahi · 2022-02-08 17:41
axboenull: with the options above, it's exactly equal to syscalls per second#asahi · 2022-02-08 17:40
axboenull: it keeps running until you ctrl-c it, unless you set a runtime#asahi · 2022-02-08 17:30
axboecompiling...#asahi · 2022-02-08 17:29
axboej`ey: yep#asahi · 2022-02-08 17:29
axboeand that's purely syscall bound#asahi · 2022-02-08 17:29
axboesingle thread, 10M for my x86 laptop, 4M for the m1#asahi · 2022-02-08 17:29
axboefwiw, I'm running t/io_uring -N1 -c1 -s1 from the fio repo#asahi · 2022-02-08 17:28
axboelet's see if it does anything#asahi · 2022-02-08 17:27
axboewhich is the PAC option? I can try without#asahi · 2022-02-08 17:26
axboenull: may indeed just be an artifact of differences in the arch, just reporting what I'm seeing from my very first tests#asahi · 2022-02-08 17:19
axboe16k#asahi · 2022-02-08 17:18
axboewith no batching, m1 is substantially slower. syscalls look more expensive#asahi · 2022-02-08 17:15
axboe~255M/sec if I include the two effiency cores#asahi · 2022-02-08 17:15
axboej`ey: since you asked, comparing just nop io_uring requests/sec between an x1 gen9 with a core i7-1185G7, x86 scores ~210M/sec and ~230M/sec on the m1 pro#asahi · 2022-02-08 16:28
axboej`ey: gotcha#asahi · 2022-02-08 16:27
axboej`ey: don't see audio here, was just looking at that#asahi · 2022-02-08 16:15
axboemps: I'm using an edited down config from thomas, works for me on m1 pro#asahi · 2022-02-07 23:46
axboej`ey: sorry, linked dropped here#asahi · 2022-02-07 23:33
axboeit was a substantial amount of time in a quick perf check#asahi · 2022-02-07 23:31
axboesven: but it does use the new batching, bonus points :)#asahi · 2022-02-07 23:30
axboeit'd just be one tags space of the full 64 tags#asahi · 2022-02-07 23:29
axboeif they are shared, then they are unique across the two queues#asahi · 2022-02-07 23:28
axboeor actually use the shared tag space support for them?#asahi · 2022-02-07 23:27
axboesven: looks like you picked 8 tags for admin queue, might make more sense to shrink that down even further#asahi · 2022-02-07 23:26
axboeirq overhead looks high on the platform#asahi · 2022-02-07 23:26
axboeit's more of a bummer that it isn't possible to experiment :)#asahi · 2022-02-07 23:26
axboenot really on a laptop, doesn't matter#asahi · 2022-02-07 23:26
axboeyeah it's very odd