File copy to local disk Windows 2019/2022 Server

[Sigi13]

Hi

We have a very strange behaviour on our Windows 2019/2022 Servers. 2016 does not seems to be affected.

We measure the copy process for a 2Gb file on a vm. I'm using the diskspd tool from Micrsofot with this parameters:

.\diskspd -t2 -o32 -b4k -r4k -w0 -d60 -Sh -D -L -c2G D:\IO.dat > test01.txt

If the VM has no snapshot, we get this:

Total IO
thread | bytes | I/Os | MiB/s | I/O per s | AvgLat | IopsStdDev | LatStdDev | file
------------------------------------------------------------------------------------------------------------------
0 | 2984321024 | 728594 | 47.43 | 12141.91 | 2.635 | 723.90 | 1.350 | D:\IO.dat (2GiB)
1 | 2995494912 | 731322 | 47.61 | 12187.37 | 2.625 | 721.81 | 1.299 | D:\IO.dat (2GiB)
------------------------------------------------------------------------------------------------------------------
total: 5979815936 | 1459916 | 95.04 | 24329.28 | 2.630 | 1426.31 | 1.325

If the same VM has snapshot, we have :

Total IO
thread | bytes | I/Os | MiB/s | I/O per s | AvgLat | IopsStdDev | LatStdDev | file
------------------------------------------------------------------------------------------------------------------
0 | 5964926976 | 1456281 | 94.79 | 24265.03 | 1.318 | 1086.55 | 0.822 | D:\IO.dat (2GiB)
1 | 5954260992 | 1453677 | 94.62 | 24221.64 | 1.320 | 1011.55 | 0.794 | D:\IO.dat (2GiB)
------------------------------------------------------------------------------------------------------------------
total: 11919187968 | 2909958 | 189.40 | 48486.66 | 1.319 | 2034.48 | 0.808

Why do we have factor 2 MiB/s(95.04 MiB/s without / 189.40 MiB/s with Snapshot) if the vm has a snapshot? Is there an explanation for that?

[kastlr]

Hi,

what type of vmdk did you create for the Windows Server 2019, thin, lazy or eagerzeroed thick?

Except for the last kind the following will happen when you would write to a never been written track.

ESXi will pause the application IO and send zeros to wipe that track before it would allow your application to proceed.

So when running copy tests creating eagerzeroed thick volumes is a must have.

Regards,

Ralf

[Sigi13]

Hi, 

All disks are thin provsioned.

I'll try it with eagerzeroed to have a comparison.

[Sigi13]

First, thank you for the tip 👍

Here are the results with a "Thick Provision Eager Zeroed" disk

with snapshot

66.97 MiB/s

without snapshot

66.44 MiB/s 

[kastlr]

Hi,

are these numbers per Thread or per Test?

I might be wrong, but per my understanding your tests are 100% reads only.

[Sigi13]

Yes this was 100% Read

100% Write looks like this from another tool. The first 4 Tests are without Snapshot. Disk 😧 is thin provisioned and E: eagerlazyzeroed

[Sigi13]

With snapshot

without