[ovirt-users] Slow conversion from VMware in 4.1

[Richard W.M. Jones]

On Tue, Feb 06, 2018 at 11:11:37AM +0100, Luca 'remix_tj' Lorenzetto wrote:
> Il 6 feb 2018 10:52 AM, "Yaniv Kaul" <ykaul at redhat.com> ha scritto:
> 
> 
> I assume its network interfaces are also a bottleneck as well. Certainly if
> they are 1g.
> Y.
> 
> 
> That's not the case, vcenter uses 10g and also all the involved hosts.
> 
> We first supposed the culprit was network, but investigations has cleared
> its position. Network usage is under 40% with 4 ongoing migrations.

The problem is two-fold and is common to all vCenter transformations:

(1) A single https connection is used and each block of data that is
requested is processed serially.

(2) vCenter has to forward each request to the ESXi hypervisor.

(1) + (2) => most time is spent waiting on the lengthy round trips for
each requested block of data.

This is why overlapping multiple parallel conversions works and
(although each conversion is just as slow) improves throughput,
because you're filling in the long idle gaps by serving other
conversions.

This is also why other methods perform so much better.  VMX over SSH
uses a single connection but connects directly to the ESXi hypervisor,
so cause (2) is eliminated.  VMX over NFS eliminates VMware servers
entirely and can make multiple parallel requests, eliminating (1) and
(2).  VDDK [in ideal circumstances] can mount the FC storage directly
on the conversion host meaning the ordinary network is not even used
and all requests travel over the SAN.

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/