By the way: Does anyone do some system profiling (flame graphs or something like that) while testing devices to make sure you reached the devices' capabilities instead of the systems? I just made some tests of nvme drives and a friend of mine (who is a scientist) said that if that was a paper he would reject it because I do not know whether I measure the devices or the system. Am 16.02.21 um 20:51 schrieb Jens Axboe: > On 2/16/21 12:00 PM, Ober, Frank wrote: >> Hi Daniel, Intel uses pvsync2 for QD1 testing so if you cannot use spdk or io_uring (these are the fastest ways to test a device), you can switch your engine to pvsync2 on this kernel and this will work. >> >> Here's what I tried on PCIe Gen3 system with an Optane drive and your drive. >> >> >> [root@fm42adsdemo001 block]# fio --bs=1M --rw=randwrite --numjobs=1 --iodepth=1 --size=10000M \ >>> --filename=/dev/nvme6n1 --name=mylittletest --direct=1 --fsync=1 \ >>> --refill_buffers --ioengine=pvsync2 --group_reporting \ >>> --fallocate=none --runtime=60 --time_based --hipri >> mylittletest: (g=0): rw=randwrite, bs=(R) 1024KiB-1024KiB, (W) 1024KiB-1024KiB, (T) 1024KiB-1024KiB, ioengine=pvsync2, iodepth=1 >> fio-3.24-5-g2ee2 >> Starting 1 process >> Jobs: 1 (f=1): [w(1)][100.0%][w=2048MiB/s][w=2048 IOPS][eta 00m:00s] >> >> I ran the same exact job on this mid-range Intel Xeon cpu and found io_uring only very slightly faster than pvsync2 >> Jobs: 1 (f=1): [w(1)][100.0%][w=2057MiB/s][w=2057 IOPS][eta 00m:00s] > That's not too surprising, if you're at queuedepth=1 or doing sync IO, then > I'd consider it a major win that the async API is just as fast or faster > than the sync one :-) > > -- Daniel Menzel Geschäftsführer Menzel IT GmbH Charlottenburger Str. 33a 13086 Berlin +49 (0) 30 / 5130 444 - 00 daniel.menzel@xxxxxxxxxxxxx https://menzel-it.net Geschäftsführer: Daniel Menzel, Josefin Menzel Unternehmenssitz: Berlin Handelsregister: Amtsgericht Charlottenburg Handelsregister-Nummer: HRB 149835 B USt-ID: DE 309 226 751
