DD Device Poor Performace on SSD/NVME

wi3dzma · Thu Nov 16, 2017 1:30 pm

Hi,

I am testing Starwind DD Device to export local ssd/nvme storage as iSCSI. Performance of Starwind Device attached on a loopback through ISCSI (so there is no network involved) is only 25% of possible performance of Storage Spaces Mirror Virtual Disk (with NTFS) that is underneath (tests done using diskspd).

Operating system: MS Hyper-V Server 2016 with latest patches

8K Random Reads (diskspd):
Storage Spaces Mirror Virtual Disk (with NTFS): 160.000 IOPS
Starwind DD Device: 40.000 IOPS

Also testing that Starwind DD Device exposed through SMB3 (40Gb/s NICs) gives even lower results: 4000 IOPS.

Any clues what is causing such low performance?

Best Regards,
Malwina Leis

Thu Nov 16, 2017 3:55 pm

How many loopback sessions do you use? I assume this was tested using one session. If so, add another loopback session (or even the third one) and test it that way. With fast SSD it would be optimum to have 3 loopback sessions. Feel free to report the results you get. Additionally, could you specify the particular interest in 8K blocks? Most applications would use 4K and 64K blocks.

wi3dzma · Thu Nov 16, 2017 4:49 pm

I had 2 sessions (on one iscsi connection). Increasing them to 3 (or to 6) didn't change the results.

I'm running multiple tests with blocks from 4K to 1M and 25% of performance is true for all of them. I have put 8K results just as an example, because this is the "Normalized IOPS" for Hyper-V.

Fri Nov 17, 2017 5:19 pm

Please submit a support ticket for better interaction and investigation of your issue and indicate a reference to this topic. We would need logs and description of the procedure of testing.

wi3dzma · Tue Nov 21, 2017 2:02 pm

Ok, I will.
I'm using diskspd to test - this is a pretty common tool to use on windows to test performance.
What performance overhead is normal for DD Device?

Best Regards,
Malwina Leis

Tue Nov 21, 2017 2:53 pm

We have got your case. Will keep communicating there.

Thu Dec 07, 2017 2:32 pm

Just to give the community an update on this case:

Testing the configuration revealed that the performance was limited by the processor used for the test server. We managed to get better results than reported by wi3dzma before, but StarWind loaded the processor up to 100% on some test patterns.