On 5/31/18 8:49 AM, Paolo Valente wrote: > > >> Il giorno 31 mag 2018, alle ore 16:38, Jens Axboe <axboe@xxxxxxxxx> ha scritto: >> >> On 5/31/18 2:55 AM, Paolo Valente wrote: >>> >>> >>>> Il giorno 27 mag 2018, alle ore 16:24, Sitsofe Wheeler <sitsofe@xxxxxxxxx> ha scritto: >>>> >>>> Hi Paolo! >>>> >>>> On 25 May 2018 at 20:20, Paolo Valente <paolo.valente@xxxxxxxxxx> wrote: >>>>> Hi, >>>>> if I run this job (even with the last version GitHub version of fio) on an SSD: >>>>> [global] >>>>> ioengine=sync >>>>> time_based=1 >>>>> runtime=20 >>>>> readwrite=randread >>>>> size=100m >>>>> numjobs=1 >>>>> invalidate=1 >>>>> [job1] >>>>> >>>>> then, after little time (I think after 100MB have been read), fio reports a nonsensically large value for the throughput, while a simple iostat shows that no I/O is going on. By just replacing time_based with loops, i.e., with a job file like this: >>>>> >>>>> [global] >>>>> ioengine=sync >>>>> loops=1000 >>>>> readwrite=randread >>>>> size=100m >>>>> numjobs=1 >>>>> invalidate=1 >>>>> [job1] >>>>> >>>>> the problem disappears. >>>> >>>> I've taken a stab at fixing this over in >>>> https://github.com/sitsofe/fio/tree/random_reinvalidate - does that >>>> solve the issue for you too? >>> >>> Nope :( >>> >>>> ... >>>> I know I'm "Teaching grandmother to suck eggs" given that you're the >>>> author of BFQ but just in case... >>>> >>>> This issue happens on loops=1000 too and I believe it's down to >>>> readahead. >>> >>> I'm afraid there is a misunderstanding on this, grandson :) >>> >>> As I wrote, this problem does not occur with loops=1000. My >>> impression is that, with loops, as well as with time-based and read, >>> fio does invalidate the cache every time it restarts reading the same >>> file, while with time_based and randread it does not (or maybe it >>> tries to, but fails for some reason). >> >> This is basically by design. loop will go through the full >> open+invalidate, whereas time_based will just keep chugging >> along. Once your 100mb is in page cache, then no more IO >> will be done, as reads are just served from there. >> > > Such a design confused me. Highlighting somewhere this deviation > (between loops and time_based) might help other dull people like me. Actually, I'm misremembering, and we did fix this up. But looks like I botched the fix, try pulling a new update and it should work for you. Fix: http://git.kernel.dk/cgit/fio/commit/?id=80f021501fda6a6244672bb89dd8221a61cee54b -- Jens Axboe -- To unsubscribe from this list: send the line "unsubscribe fio" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
