[Libguestfs] nbdkit, VDDK, extents, readahead, etc

[Richard W.M. Jones]

As I've spent really too long today investigating this, I want to
document this in a public email, even though there's nothing really
that interesting here.  One thing you find from search for VDD 6.7 /
VixDiskLib_QueryAllocatedBlocks issues with Google is that we must be
one of the very few users out there.  And the other thing is that it's
quite broken.

All testing was done using two baremetal servers connected back to
back through a gigabit ethernet switch.  I used upstream qemu and
nbdkit from git as of today.  I used a single test Fedora guest with a
16G thin provisioned disk with about 1.6G allocated.

Observations:

(1) VDDK hangs for a really long time when using the nbdkit --run
    option.

It specifically hangs for exactly 120 seconds doing:

  nbdkit: debug: VixDiskLib: Resolve host.

This seems to be a bug in VDDK, possibly connected with the fact that
we fork after initializing VDDK but before doing the
VixDiskLib_ConnectEx.  I suspect it's something to do with the PID
changing.

It would be fair to deduct 2 minutes from all timings below.

(2) VDDK cannot use VixDiskLib_QueryAllocatedBlocks if the disk is
opened for writes.  It fails with this uninformative error:

  nbdkit: vddk[1]: error: [NFC ERROR] NfcFssrvrProcessErrorMsg: received NFC error 13 from server: NfcFssrvrOpen: Failed to open '[datastore1] Fedora 28/Fedora 28.vmdk'
  nbdkit: vddk[1]: error: [NFC ERROR] NfcFssrvrClientOpen: received unexpected message 4 from server
  nbdkit: vddk[1]: debug: VixDiskLib: Detected DiskLib error 290 (NBD_ERR_GENERIC).
  nbdkit: vddk[1]: debug: VixDiskLib: VixDiskLibQueryBlockList: Fail to start query process. Error 1 (Unknown error) (DiskLib error 290: NBD_ERR_GENERIC) at 543.
  nbdkit: vddk[1]: debug: can_extents: VixDiskLib_QueryAllocatedBlocks test failed, extents support will be disabled: original error: Unknown error

The last debug statement is from nbdkit itself indicating that because
VixDiskLib_QueryAllocatedBlocks didn't work, extents support is
disabled.

To work around this you can use nbdkit --readonly.  However I don't
understand why that would be necessary, except perhaps it's just an
undocumented limitation of VDDK.  For all the cases _we_ care about
we're using --readonly, so that's lucky.

(3) Using nbdkit-noextents-filter and nbdkit-stats-filter we can
nicely measure the benefits of extents:

With noextents (ie. force full copy):

  elapsed time: 323.815 s
  read: 8194 ops, 17179869696 bytes, 4.24437e+08 bits/s

Without noextents (ie. rely on qemu-img skipping sparse bits):

  elapsed time: 237.41 s
  read: 833 ops, 1734345216 bytes, 5.84423e+07 bits/s
  extents: 70 ops, 135654246400 bytes, 4.57114e+09 bits/s

Note if you deduct 120 seconds (see point (1) above) from these times
then it goes from 203s -> 117s, about a 40% saving.  We can likely do
better by having > 32 bit requests and qemu not using
NBD_CMD_FLAG_REQ_ONE.

(4) We can also add nbdkit-readahead-filter in both cases to see if
that helps or not:

With noextents and readahead:

  elapsed time: 325.358 s
  read: 265 ops, 17179869184 bytes, 4.22423e+08 bits/s

As expected the readahead filter reduces the numbers of iops greatly.
But in this back-to-back configuration VDDK requests are relatively
cheap so no time is saved.

Without noextents, with readahead:

  elapsed time: 252.608 s
  read: 96 ops, 1927282688 bytes, 6.10363e+07 bits/s
  extents: 70 ops, 135654246400 bytes, 4.29612e+09 bits/s

Readahead is detrimental in this case, as expected because this filter
works best when reads are purely sequential, and if not it will tend
to prefetch extra data.  Notice that the number of bytes read is
larger here than in the earlier test.

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
virt-df lists disk usage of guests without needing to install any
software inside the virtual machine.  Supports Linux and Windows.
http://people.redhat.com/~rjones/virt-df/