#asahi-dev · 2022-02-19 08:28
marcan*those#asahi-dev · 2022-02-19 08:28
marcanjannau: ah, thanks, missed thsoe#asahi-dev · 2022-02-19 08:28
marcan(which is what j`ey said)#asahi-dev · 2022-02-19 08:28
marcanpovik: I pushed asahi-soc/next, that#asahi-dev · 2022-02-19 08:25
marcanah right, that bug#asahi-dev · 2022-02-19 08:24
marcanpovik: any new audio updates?#asahi-dev · 2022-02-19 08:23
marcansven: nvme stuff? I think RTKit is already in order in the spmi branch, right?#asahi-dev · 2022-02-19 08:23
marcanam I forgetting anything other than what was already in spmi/work and jannau's spi fixes?#asahi-dev · 2022-02-19 08:22
marcanalright, so things for the next asahi#asahi-dev · 2022-02-19 07:46
marcanoh, well then yeah#asahi-dev · 2022-02-19 07:45
marcannot sure what test sven was doing but I'm pretty sure you aren't flushing each write that way if you're getting that performance#asahi-dev · 2022-02-19 07:43
marcanno --fsync?#asahi-dev · 2022-02-19 07:42
marcanand it's still slow because it still has to flush once in a while, but can probably enforce barrier semantics across multiple writes through the metadata without flushing every time#asahi-dev · 2022-02-19 07:41
marcanthe barrier flush thing in macos is probably just something it constructs on top of the cow/journal of the filesystems#asahi-dev · 2022-02-19 07:41
marcanI don't think drive barriers are a thing#asahi-dev · 2022-02-19 07:40
marcanrkjnsn: that's partially a filesystem thing; according to axboe linux doesn't really do barriers these days#asahi-dev · 2022-02-18 07:10
marcanoh, jannau already mentioned it#asahi-dev · 2022-02-18 07:10
marcanaxboe: wait, shouldn't that be F_FULL*F*SYNC on the patch? did you compile test it? :-)#asahi-dev · 2022-02-18 05:41
marcanaxboe: nice :)#asahi-dev · 2022-02-18 05:40
marcanI got 350 on an M1 Max lol#asahi-dev · 2022-02-18 05:40
marcandepending on the machine, apparently?#asahi-dev · 2022-02-18 05:40
marcanrkjnsn: it does 300-600 IOPS, which is still slow#asahi-dev · 2022-02-17 14:57
marcanbut fsync() certainly wouldn't save macOS#asahi-dev · 2022-02-17 14:56
marcanso it might just be unsafe on any system :p#asahi-dev · 2022-02-17 14:56
marcanthat said, I'm not even entirely sure if rsync tries to do a normal fsync() here#asahi-dev · 2022-02-17 14:48
marcanhow's that for a scary demo?#asahi-dev · 2022-02-17 14:48
marcanhttps://news.ycombinator.com/item?id=30373572#asahi-dev · 2022-02-17 14:35
marcanI did set that for notifications :p#asahi-dev · 2022-02-17 14:14
marcanthe only reason I haven't deleted twitter is those 50k followers are *sometimes* useful :p#asahi-dev · 2022-02-17 14:04
marcansince it doesn't matter on iOS#asahi-dev · 2022-02-17 14:04
marcanI suspect nobody noticed#asahi-dev · 2022-02-17 13:45
marcan:p#asahi-dev · 2022-02-17 13:44
marcanalyssa: yes, you get better database transaction latencies on a shitty flash drive than on internal NVMe on these machines :-)#asahi-dev · 2022-02-17 13:44
marcan(of course, the flash drive has no cache at all, so it's equally slow with a vanilla fsync :p)#asahi-dev · 2022-02-17 13:43
marcanfor shits and giggles (and because HN commenters are tiring): full fsync on a shitty USB3 flash drive on macOS/M1: 223 IOPS. internal NVMe, 58 IOPS. Both FAT32.#asahi-dev · 2022-02-17 10:16
marcanI wonder if it, like, linearly scans some huge cache hashtable?#asahi-dev · 2022-02-17 10:16
marcanso yeah, that NVMe sync command is making ANS2 do a lot of work...#asahi-dev · 2022-02-17 10:16
marcanthat's more like it#asahi-dev · 2022-02-17 10:16
marcanANS2 DCS WR : 0.411 MB/s#asahi-dev · 2022-02-17 10:16
marcanANS2 DCS RD : 0.244 MB/s#asahi-dev · 2022-02-17 10:16
marcansame test without the full fsync: write: 93.52 ops/s 397.77 KBytes/s#asahi-dev · 2022-02-17 10:15
marcanso yeah, ANS2 is definitely doing something somewhat dodgy if it's doing 9MB/s of memory write traffic and 3MB/s of memory read traffic to serve 320KB/s of data traffic#asahi-dev · 2022-02-17 10:14
marcanANS2 WR : 8.836 MB/s#asahi-dev · 2022-02-17 10:14
marcanANS2 RD : 3.042 MB/s#asahi-dev · 2022-02-17 10:14
marcanand:#asahi-dev · 2022-02-17 10:14
marcanon FAT32 it's actually slower (34 IOPS), but no serious amplification: 78.37 ops/s 321.02 KBytes/s#asahi-dev · 2022-02-17 10:11
marcanlet me try another filesystem...#asahi-dev · 2022-02-17 10:11
marcanso that sounds like two problems here... FULLFSYNC does some horrible APFS amplification nonsense, *and* ANS2 blows it up even more#asahi-dev · 2022-02-17 10:10
marcandoing the same thing, but with an artificial delay to make it run at the same speed sans the flush, I get 44 disk ops/s / 180 KB/s, and ~200 KB/s of ANS2 memory bandwidth#asahi-dev · 2022-02-17 10:09
marcanwhile running the python loop, python says 46 ops/s, powermetrics reports 915 disk ops per second / 4.7MB/s (seems a bit much? APFS write amplification? and 20MB/s of ANS2 memory bandwidth in both directions)