#asahi-dev · 2022-02-16 19:54
axboeyeah that looks inline with what I see#asahi-dev · 2022-02-16 19:54
axboeif it created them#asahi-dev · 2022-02-16 19:54
axboeyou can add --unlink=1, then fio will remove the test file(s) after the run#asahi-dev · 2022-02-16 19:53
axboeI get ~12k first pass, ~50k subsequent pass#asahi-dev · 2022-02-16 19:52
axboemaybe rm test file and see if it reproduces#asahi-dev · 2022-02-16 19:52
axboeis it an overwrite pass?#asahi-dev · 2022-02-16 19:51
axboemaybe your drive is faster than mine?#asahi-dev · 2022-02-16 19:49
axboedata vs data + metadata#asahi-dev · 2022-02-16 19:48
axboeit's similar to fdatasync#asahi-dev · 2022-02-16 19:46
axboebut not touch disk#asahi-dev · 2022-02-16 19:46
axboeas otherwise the job may just be dirtying page cache, and run very fast#asahi-dev · 2022-02-16 19:46
axboeGlanzmann: end_fsync will make a potentially big difference if you're just doing regular buffered writes#asahi-dev · 2022-02-16 19:46
axboeif you change sync=1 to sync=dsync, for example, I get ~50K IOPS#asahi-dev · 2022-02-16 19:44
axboeso won't change much#asahi-dev · 2022-02-16 19:44
axboeGlanzmann: end_fsync is just an fsync() when all 1G have been written#asahi-dev · 2022-02-16 19:43
axboemight be differences in what O_SYNC provides?#asahi-dev · 2022-02-16 19:43
axboeso it's not the flushes at that point#asahi-dev · 2022-02-16 19:43
axboeeven doing no flushes on linux will be about the same as 1 per second#asahi-dev · 2022-02-16 19:43
axboeGlanzmann: not right now, no#asahi-dev · 2022-02-16 19:39
axboeit really should be done at the nvme_ns level, or on the block side. but... would rather keep it local#asahi-dev · 2022-02-16 19:37
axboea 200x speedup at minimal risk warrants a hack ;)#asahi-dev · 2022-02-16 19:35
axboesven putting on a polite face ;)#asahi-dev · 2022-02-16 19:24
axboeGlanzmann: let me know if you try the above and if there's any noticeable difference between that and using the write through hack#asahi-dev · 2022-02-16 19:24
axboehack alert...#asahi-dev · 2022-02-16 19:24
axboehttps://git.kernel.dk/cgit/linux-block/commit/?h=m1-test&id=50c69de75b60f6e47dc9f2a2ee85c6f69648e489#asahi-dev · 2022-02-16 18:21
axboelet me hack it up#asahi-dev · 2022-02-16 18:21
axboeyeah, might be useful to just benchmark for typical workloads#asahi-dev · 2022-02-16 18:20
axboebut for this one, probably a 5 second period for flush would be useful#asahi-dev · 2022-02-16 18:20
axboeoh yeah, LD_PRELOAD hack was popular#asahi-dev · 2022-02-16 18:19
axboeside bar, I actually wrote laptop_mode back in the day#asahi-dev · 2022-02-16 18:19
axboemarcan: no, that was on dirty data writeback#asahi-dev · 2022-02-16 18:18
axboeGlanzmann: yeah, I noticed that too, just flies through the install part#asahi-dev · 2022-02-16 18:13
axboekinda nasty, but safer#asahi-dev · 2022-02-16 18:13
axboe"only issue flush if X seconds has passed since last issue"#asahi-dev · 2022-02-16 18:12
axboealternatively, some sort of time based hack might make sense#asahi-dev · 2022-02-16 18:12
axboeGlanzmann: it'll bump your test case from 56 iops to 14k or something like that :)#asahi-dev · 2022-02-16 18:12
axboeGlanzmann: and yes, that's what it means#asahi-dev · 2022-02-16 18:11
axboeGlanzmann: ^^#asahi-dev · 2022-02-16 18:11
axboeI just do: `echo "write through" > /sys/block/nvme0n1/queue/write_cache"` for now...#asahi-dev · 2022-02-16 18:06
axboemarcan: if it's guaranteed by the controller itself, it should just not advertise a volatile write cache#asahi-dev · 2022-02-16 18:05
axboemarcan: was thinking that's probably the case on osx, and hence nobody ever cared about flush latencies#asahi-dev · 2022-02-16 17:38
axboeoutside of missing some weird init and not having timings for osx, guess I'm just going to mark it 'write through' for now :/#asahi-dev · 2022-02-16 17:22
axboeand even if there are no writes, it's still a 20-25 usec request#asahi-dev · 2022-02-16 17:22
axboeon the data I collected, nothing else seems useful apart from the fact that writes-since-flush is always != for a slow flush, and only fast if writes-since-flush == 0#asahi-dev · 2022-02-16 17:21
axboeso it could very well be that the device just sucks at flushes and we're screwed#asahi-dev · 2022-02-16 17:21
axboesven: heh ok, curious#asahi-dev · 2022-02-16 17:20
axboesven: have you looked at osx timings? I seem to recall you said it just does fewer flushes, but are all flushes slow there too#asahi-dev · 2022-02-16 17:20
axboe16.5 -> 17.5 msec#asahi-dev · 2022-02-16 17:19
axboeeven if we only did a single write, and that write was FUA, the flush is still slow#asahi-dev · 2022-02-16 17:19
axboewhich is unsurprising