I've tried fio directly on nvme device and without filesystem. The
drop with fsync is not that significant; 44313 vs 42713 on a 30s
randwrite run with iodepth=1


# fio -filename=/dev/nvme0n1 -direct=1 -ioengine=libaio -rw=randwrite
-bs=4k -name=mytest -iodepth=1 -runtime=30 -time_based
mytest: (g=0): rw=randwrite, bs=4K-4K/4K-4K/4K-4K, ioengine=libaio, iodepth=1
fio-2.1.3
Starting 1 process
Jobs: 1 (f=1): [w] [100.0% done] [0KB/177.6MB/0KB /s] [0/45.5K/0 iops]
[eta 00m:00s]
mytest: (groupid=0, jobs=1): err= 0: pid=2131: Thu Feb  9 18:56:01 2017
  write: io=5193.2MB, bw=177253KB/s, iops=44313, runt= 30001msec

# fio -filename=/dev/nvme0n1 -direct=1 -ioengine=libaio -rw=randwrite
-bs=4k -name=mytest -iodepth=1 -runtime=30 -time_based -fsync=1
mytest: (g=0): rw=randwrite, bs=4K-4K/4K-4K/4K-4K, ioengine=libaio, iodepth=1
fio-2.1.3
Starting 1 process
Jobs: 1 (f=1): [w] [100.0% done] [0KB/167.4MB/0KB /s] [0/42.9K/0 iops]
[eta 00m:00s]
mytest: (groupid=0, jobs=1): err= 0: pid=2136: Thu Feb  9 19:04:54 2017
  write: io=5005.5MB, bw=170853KB/s, iops=42713, runt= 30000msec


On Wed, Feb 22, 2017 at 3:40 AM, Vojtech Pavlik [off-list ref] wrote:

On Tue, Feb 21, 2017 at 10:48:06AM -0600, shiva rkreddy wrote:

quoted

fio command without fsync:

# fio -filename=/dev/bcache0 -direct=1 -ioengine=libaio -rw=randwrite
-bs=4k -name=mytest -iodepth=1 -runtime=30 -time_based

iops : 35k

fio command with fsync:

fio -filename=/dev/bcache0 -direct=1 -ioengine=libaio -rw=randwrite
-bs=4k -name=mytest -iodepth=1 -runtime=30 -time_based -fsync=1

iops: 8.1k

quoted

I'm quite surprised by the drop in iops with fsync turned on. Is this
expected or am I missing some basic setting?

It's not uncommon that fsync would have a huge performance impact.
Without fsync, most of the data never hits the storage and is only
staying in the system memory.

May I suggest that you try to measure the performance of the same tests
when the filesystem is created on the NVMe device directly, without
using bcache? You're likely to observe a similar pattern.

--
Vojtech Pavlik
Director SUSE Labs