On Fri, 9 Nov 2018 at 00:18, Jens Axboe <axboe@xxxxxxxxx> wrote: > > On 11/8/18 1:09 PM, Sebastien Boisvert wrote: > > > > > > On 2018-11-07 3:57 p.m., Jens Axboe wrote: > >> On 11/5/18 8:00 AM, Sebastien Boisvert wrote: > >>> > >>> > >>> On 2018-11-04 5:44 a.m., Sitsofe Wheeler wrote: > >>>> Looks like someone is referencing an fio benchmark result on Apple's > >>>> Mac Mini page and whoever did it took care to respect the Moral > >>>> License (https://fio.readthedocs.io/en/latest/fio_doc.html#moral-license > >>>> ). From https://www.apple.com/mac-mini/ : > >>>> > >>>> "4. Testing conducted by Apple in October 2018 using preproduction > >>>> 3.2GHz 6-core Intel Core i7-based Mac mini systems with 64GB of RAM > >>>> and 1TB SSD, and shipping 3.0GHz dual-core Intel Core i7-based Mac > >>>> mini systems with 16GB of RAM and 1TB SSD. Tested with FIO 3.8, 1024KB > >>>> request size, 150GB test file and IO depth=8. Performance tests are > >>>> conducted using specific computer systems and reflect the approximate > >>>> performance of Mac mini." > >>>> > >>>> My only question is: as the depth was 8 were they using the posixaio engine? > >>>> > >>> > >>> The foot note number 4 supports this claim: > >>> > >>> "Up to 4X faster read speed" > >>> > >>> It would make sense to use asynchronous I/O since ioengine=psync is the default on Mac. > >> > >> I'd be fine making that change, if someone can benchmark psync vs posixaio > >> in terms of latency in that platform. > >> > >> Might also make sense to improve the setup so that we have a default > >> engine per OS depending on iodepth. For instance, on Linux, QD=1 should > >> just be psync. But if QD > 1, then we should default to libaio. I'm > >> afraid lots of folks have run iodepth=32 or whatever without changing > >> the IO engine and wondering what is going on. > > > > Would this change be *after* parse_options() has been called ? > > > > I looked at init.c and options.{h,c}. > > Right now we use FIO_PREFERRED_ENGINE to set the default engine, which > can be defined by the platform. I think we should drop that, don't set > a default, and instead add some logic to eg fixup_options() that sets > the preferred engine based on platform and depth. Probably want > platforms to define > > FIO_PREF_SYNC_ENGINE > FIO_PREF_ASYNC_ENGINE > > and just pick one of those depending on iodepth. Something like that. I don't know this is a safe idea unless we are also going to start tweaking other defaults at the same time. Imagine someone going from iodepth=1 to iodepth=2 but with direct=0 set - ihey are now comparing pvsync iodepth=1 to libaio iodepth=2. Perhaps there should be an "best" ioengine that does what described? I still wonder whether we should just warn if someone uses a synchronous engine with an iodepth > 1 when they aren't using one of the typical cases (https://github.com/axboe/fio/pull/347 + (libaio with direct=1) ) ... -- Sitsofe | http://sucs.org/~sits/
