On 01/08/2016 12:35 PM, Jeff Furlong wrote:
Good points. Here is further data: # blockdev --getsize64 /dev/nvme2n1 800166076416 # fio --name=SW_1MB_QD32 --ioengine=libaio --direct=1 --rw=write --iodepth=32 --size=1000m --runtime=60s --time_based --numjobs=1 --bs=1m --overwrite=1 --filename=/dev/nvme2n1 SW_1MB_QD32: (g=0): rw=write, bs=1M-1M/1M-1M/1M-1M, ioengine=libaio, iodepth=32 fio-2.3-11-g5f3b Starting 1 process Jobs: 1 (f=1): [W(1)] [100.0% done] [0KB/990.2MB/0KB /s] [0/990/0 iops] [eta 00m:00s] SW_1MB_QD32: (groupid=0, jobs=1): err= 0: pid=19511: Fri Jan 8 11:08:55 2016 write: io=58620MB, bw=976.11MB/s, iops=976, runt= 60001msec So the above using a fixed --size that is compatible with --bs is one workaround. When trying to write more than 100% of the device (wrapping around after end of device), instead of --size=100% we could use a fixed size (where above 800166076416 bytes is 763097.8 MB), even if time_based is not used (use --io_size instead): # fio --name=SW_1MB_QD32 --ioengine=libaio --direct=1 --rw=write --iodepth=32 --size=763097m --io_size=764000m --numjobs=1 --bs=1m --overwrite=1 --filename=/dev/nvme2n1 SW_1MB_QD32: (g=0): rw=write, bs=1M-1M/1M-1M/1M-1M, ioengine=libaio, iodepth=32 fio-2.3-11-g5f3b Starting 1 process Jobs: 1 (f=1): [W(1)] [100.0% done] [0KB/567.9MB/0KB /s] [0/567/0 iops] [eta 00m:00s] SW_1MB_QD32: (groupid=0, jobs=1): err= 0: pid=129928: Thu Jan 7 12:45:52 2016 write: io=764000MB, bw=600510KB/s, iops=586, runt=1302786msec Alternatively, if we simply back out the commit c82ea3d49aa then we can make the 1% or 100% options (--time_based or --io_size) work fine: # fio --name=SW_1MB_QD32 --ioengine=libaio --direct=1 --rw=write --iodepth=32 --size=1% --runtime=60s --time_based --numjobs=1 --bs=1m --overwrite=1 --filename=/dev/nvme2n1 SW_1MB_QD32: (g=0): rw=write, bs=1M-1M/1M-1M/1M-1M, ioengine=libaio, iodepth=32 fio-2.3-11-g5f3b Starting 1 process Jobs: 1 (f=1): [W(1)] [100.0% done] [0KB/954.2MB/0KB /s] [0/954/0 iops] [eta 00m:00s] SW_1MB_QD32: (groupid=0, jobs=1): err= 0: pid=21853: Fri Jan 8 11:20:04 2016 write: io=56931MB, bw=971606KB/s, iops=948, runt= 60001msec
So the bug isn't really in the bisected commit, it just changes how fio backs out and then shows another bug. The real bug is that we don't align the new start properly, so we end up with these weird unaligned start offsets that don't work with O_DIRECT IO.
Try current -git, I just committed this fix: http://git.kernel.dk/cgit/fio/commit/?id=19ddc35b9b97be5af371bb65e93a4864d1dce7b6 -- 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
