1

For more than a week, I am trying to determine the reason for the following IO performance degradation between proxmox host and a Windows Server 2019 VM(s).

I have to ask for your help guys because I've run out of ideas.

Environment data:

  • Single proxmox host, no cluster, pve 6.1-8 with ZFS
  • A few WS19 VMs all having this issue, very low load, SOHO-usage
  • ZFS sync=disabled, volblocksize for VM disks = 4k
  • VM has all the latest VirtIO drivers (0.1.173)

IO test on both VM and host with the following fio command:

fio --filename=test --sync=1 --rw=$TYPE --bs=$BLOCKSIZE --numjobs=1 --iodepth=4 --group_reporting --name=test --filesize=1G --runtime=30

Results (Host vs VM):

  • 4K Random Read: 573 vs 62.5 MiB/s
  • 4K Random Write: 131 vs 14.1 MiB/s
  • 4K Sequential Read: 793 vs 56,2 MiB/s
  • 4K Sequential Write: 240 vs 3,42 MiB/s
  • 64K Random Read: 1508 vs 831 MiB/s
  • 64K Random Write: 596 vs 62,5 MiB/s
  • 64K Sequential Read: 1631 vs 547 MiB/s
  • 64K Sequential Write: 698 vs 43,8 MiB/s

Charts:

Results Table

Reads Chart

Writes Chart

What I have tried so far: different volblocksizes on ZFS, different ZFS sync setting (left it on disabled, since the host is in DC), virtio-blk vs virtio scsi single (not much difference), writeback cache (became even worse).

Any suggestions what am I missing?

Programmierus
  • 11
  • 2
  • Are things more comparable if you `end_fsync=1` rather than `sync=1` at the end? Are you aware of https://github.com/axboe/fio/issues/833 (basically Windows doesn't preallocate files in quite the same way as Linux). Then again the default ioengine on Linux is `psync` (which is NOT asynchronous) so these jobs are likely not comparable across platforms anyway.. – Anon Apr 10 '20 at 05:19
  • Nice catch @Anon, thank you! With `end_fsync=1` results became truly more comparable, but the degradation of more than 3x times persists. `overwrite=1` made no difference at all though, so I don't think I am affected by #833. I am still interested to understand the roots of degradation. – Programmierus Apr 10 '20 at 14:01
  • Hmm. Judging from other "My VM is slower than my host" posts this is going to be tough to track down. Something that helps a bit is posting the output of each of the fio runs. On Linux this says how "busy" the disks were which helps... – Anon Apr 11 '20 at 14:04

0 Answers0