Server Products
Data Center Products including boards, integrated systems, Intel® Xeon® Processors, RAID Storage, and Intel® Xeon® Processors
4761 Discussions

RMS25CB080 performance

MVilf
Beginner
2,036 Views

Hi !

We have server R2208ip4lhpc (fw 02.03.0003) with RMS25CB080 (fw 23.29.0-0016), AXXRMFBU2 (fw 0016).

Drives 5x HGST HUC109060CSS600 (600gb Enterprise SAS 10k 6gbit/s) configured as two RAID1 sets and global hot-spare.

Operating system Windows 2012R2, only with Hyper-V role.

Storage driver version 14.05.02.03.

Raid module setting (on both virtual drives): strip size 256k (default), Disk cache policy: enable, Read policy: Always read ahead, IO policy: cached IO, Default write policy: Write back with BBU (current write policy= write back). (Those are the settings giving best performance.)

Problem is slow performance with reading or writing small blocks up to 64k (picture bellow).

This is (my baseline) performance from HP DL380g8 with roughly the same configuration (P420i, raid1 with 10k sas drives...)

Am i doing something wrong or is this normal?

Best regards

Marko

0 Kudos
10 Replies
MVilf
Beginner
819 Views

Hi !

I haven't read the release notes until now, but i just found out, that there were some performance issues.. They should be fixed according to the release notes of the driver.

Please Intel give me few comments.

Best regards

Marko

0 Kudos
David_A_Intel
Moderator
819 Views

Let me take a look into this. I will try to set this configuration in our lab to compare results. I will post back as soon as I have an update.

0 Kudos
MVilf
Beginner
819 Views

Hi David!

I have to let you know about some more facts regarding this issue.

- Both servers (R2200 and HPdl380g8) are running in production enviroment.

- Tests were made on off peak hours on physical hardware, when servers were more or less idle (that was the best i was able to do at the moment to trace issue further)

- We started to trace the whole issue when one older file sharing application was moved from physical hardware (Intel S5000, sata disks) to virtual machine on R2200 platform and not performing as it was on older platform.

- If I compare disk benchmarks from old sata S5000 and new virtual machine on R2200 (gen2, 4cpu, 4gb ram, fixed size scsi vhdx) the interesting thing is that old sata sistem performs better with small blocks and the virtual machine performs much better with large blocks. It is some strange nonlinearity if i compare both benchmarks. You can also observe it on physical hardware, (from the pictures i posted), but is is less apparent.

Best regards

Marko

0 Kudos
MVilf
Beginner
819 Views
0 Kudos
David_A_Intel
Moderator
819 Views

Let me work on an update for this.

0 Kudos
David_A_Intel
Moderator
819 Views

markov

Have you tried to test the drives outside of the Virtual Machine? I am trying to see if this performance impact is only within the VMs.

0 Kudos
MVilf
Beginner
819 Views

Hi !

The tests posted above (in the first post) were made directly on host running Windows 2012 R2 Standard (Hyper-V root partition).

The performance diference inside VM (Hyper-V child partition) is even worse, but only when reading/writing small blocks (attached picture). This is the benchmark from the VM that i mentioned in the post from Dec.2. It is running on the R2200 (gen2, 4cpu, 4gb ram, fixed size scsi vhdx):

As you can see, the performance in this VM shows more nonlinearity if compared to Hyper-V root partition (first post, first picture). Small blocks reads and writes are (in my opinion at least) much slower than expected..

0 Kudos
MVilf
Beginner
819 Views

Hi David !

Plese understand, that this is important for us. I realy don't want to be rude, but can you please give me some usable reply on this or direct me to someone.

0 Kudos
David_A_Intel
Moderator
819 Views

I totally understand your position. I would recommend contacting our http://www.intel.com/p/en_US/support/contactsupport Intel® Customer Support to properly follow up your concern.

0 Kudos
MVilf
Beginner
819 Views

Thanks David

I have just created support request.

0 Kudos
Reply