On Wed, Aug 22, 2018 at 04:02:42PM +0200, Martin LEUSCH wrote:
what happens with dd if=/dev/zero of=/var/srv/nfs/testfile bs=64k count=16k conv=fsync And this is direct on the server, not via NFS, right?I've got same result as previous test. tests with dd command are executed directly on the server.
Can you rebuild the partition? If so, unmount it then perform the dd directly to the device. (of=/dev/sda2 or whatever) What's the hardware raid vendor? Honestly, the performance of ext4 partition is also horrible, just not as bad as the xfs partition. What's the hardware raid model? Does it have a battery backed cache? (I will guess not, because the test is only 1G and that should fit into most current caches?) If the test directly to the partition shows good performance, then it might be an alignment issue; try recreating the fs with `mkfs.xfs -d su=64k,sw=2 ...`. If the performance is the same to the raw partition (and still much worse than the root partition) then the partition alignment itself might be off--which means recreating the partitions. (fsck and mkfs.xfs normally handle this automatically, but that depends on the raid controller+driver passing the information up, and per your xfsinfo output, that didn't happen.)
Mike Stone