Am 22.02.2017 um 10:05 hat Jaze Lee geschrieben: > I test on cache=none with sata disk, because i do not have ssd in my pc, > and I do not find any performance improved. > > I do not know why the performance is not be improved....
Most likely your disk is just too slow for the memory copy to make any significant difference, and the memcpy cost is dwarved by the actual disk I/O cost. Have you tried backing the server with a ramdisk instead? > The test case is: > boot vm from rbd, and attatch one rbd volume with 100G size to it. The > qemu and the osd is on the same host. > > The fio is > [global] > ioengine=libaio > filename=/dev/vdb > thread > group_reporting > direct=1 > iodepth=32 > > [write_32k] > bs=32k > rw=write > runtime=30 > stonewall > > [rand_write_4k] > bs=4k > rw=randwrite > runtime=30 > stonewall At least the second one is (due to its small block size) very unlikely to produce real vectored I/O, i.e. a qiov with more than just a single entry. Of course, you still get rid of a memcpy, but every request comes with additional overhead in other places, so I expect that with larger block sizes, the effect of zero-copy will be more visible. Kevin
