Testing Tintri VMStore for VIEW Implementation

[jfierberg]

Has anyone tested out the VMStore from Tintri for use in a VIEW5 environment?  We are beginning a POC with them now and want to stress the heck out of it with 30 linked clone XP desktops.  Any suggestions?  Also what has been your experience with this storage vendor?

[mittim12]

We recently moved our 200 desktop View environment over to TIntri.  I have to be honest I was a little hesitant just because I wasn't really familiar with NFS since we have always used iSCSI and FC.      During our POC I was extremely pleased with the minimal amount of effort it took to setup and deploy the infrastructure.  I was also incredible impressed with their support and our SE and they ended up being one of the reasons we went with the product.

On a performance standpoint we were honestly unable to really strain the VMstore instead we ended up stressing our storage NICS which were only 1 GB.   As a matter of fact the only downside I could find is the load balancing of NFS over the nics but that doesn't really have anything to do with Tintri.

[jfierberg]

What tools did you use with your POC to test with?

[mittim12]

We used an evaluation of http://www.loginvsi.com/ and also experimented with boot storms and linked clone provisioning. 

[jfierberg]

How have the 200 VMs on the VMstore fared compared to your other VDI systems on other storage platforms?  Could the unit handle more load?

[mittim12]

Well the VDI infrastructure was sharing the SAN with our server infrastructure so have a dedicated environment is definitively better.   We have not had any issues or slowdowns that we experienced at certain times.  I honestly feel like we could double the environment and not run into any issues at all from the storage standpoint.    I know from discussing our concerns with Tintri that they have at least one customer running around 500 desktops with no problems.   

[cgrubbe]

I have a VMStore running about 200VM's in production at this point (Linked Clones and Replicas) without any issues. 

There were a couple of different ways to stress test the box. If you are looking to max out the IOPS for the box you can use a tool like iometer with a bunch of worker processes on a handful of VM's to just generate load. There is a really good whitepaper that talks about IOPS in Windows 7 that also covers ways to test your system.  I also found a utility called VDI Sizing Tool that runs through a typical user workload (Office Apps, IE, etc.) that worked pretty well and is free. 

We have been very pleased with how our VMStore has handled every load we could imagine and so far I haven't found any real world scenario that will slow it down enough to impact the user experience.  I have launched a virus scan on 350 VM's at once which maxed out my hosts before it maxed out the VMStore (at about 65,000 IOPS - yes this was intentional).

One of the big problems we ran into with the other storage we use for our View implementation was that it couldn't support having the VM's refresh on logout.  If enough people logged out at the same time, the refresh would slow everything to a crawl.  With the VMStore the only indcation that the load has gone up is on the dashboard.

-Chris

[jfierberg]

Thank you for your response. We’ve discovered that on another storage platform, 7500 write IOPS( 8 to 64K size) slows everything to a crawl. What was your IOPS size.

[jfierberg]

Also, when you did the AV scan, did it include an definitions update?  At the moment we notice that when we do a def update to our VDI systems, we flood the NVRAM on the controller.  We are hoping that the Tintri unit handles this better.

[jslabaugh]

I have a Tintri T540 as well, and am curious to know whether your Tintri was connected with 10gig or 1gig at the time of the A/V scan.

[jfierberg]

We are into our POC now and with 30 linked clone XP vms running an IOMETER job ( 2 workers with 16KB IOP size, 80/20 write to reads), we are getting an average of 20 second write latency. Not a good stat. The Tintri unit is performing well but I am not sure what the perceived user experience is.

[mittim12]

20 second or 20 ms write latency?    Are you seeing this number fro IOMETER, RESXTOP, or the Tintri management console?

[jfierberg]

20 ms in the vm performance tab, under virtual disk. ESXTOP shows the storage unit having a ~13 in GAVG which should be below 10. Tintri’s console seems to show that the host and storage have latency too.

[jslabaugh]

Interesting. I'm runnning almost 100 production VM's on mine, mostly linked clones, and havent seen anything like those numbers. We hover around 0.5-1ms latency for the same metric, if that.

[mittim12]

This might be a good opportunity to involve their support.   That way you have more eyes looking at it and you get an opportunity to judge how support does in a real world troubleshooting scenario.

[jfierberg]

I am open to the possibility that the IO test I am running may not be correct but so far the test seems valid.

I am using the ESG-LAB-Workloads-v14 in IOMETER to simulate a virtual desktop workload.

[jfierberg]

Good point. I am working with the SE who helped me set up the unit.

[cgrubbe]

The AV scan I ran did not include def updates, but I can say that we have had issues with def updates on other storage that I haven't seen at all on the VMStore.  Our vmstore is connected using 10Gb and our hosts are also running 10Gb shared between storage, vm's and the management network.

I'm out of the office right now, so I don't have a latency stat to compare to yours, but I will say that 20ms spike isn't bad as long as it's only a spike.  Also the amount of latency you have will be based on how much throughput you have between your hosts and the VMStore.  There can be differences between the latency you see per vm, per host and from the Tintri console.

-Chris

[mittim12]

I agree with Chris in that a spike of 20 ms is not to big of a concern but a sustained average of 20 ms would not be good at all.   As far as your work load if it simulates the IO profile of your VDI environment than I think it's fine. 

[jfierberg]

We changed up the testing parameters in IOMETER to match what is in Jim Polley’s VDI Deepdive (http://jimmoyle.com/wordpress/wp-content/uploads/downloads/2011/05/Windows_7_IOPS_for_VDI_a_Deep_Dive_1_0.pdf) and got significantly better test results.

4 workers @ 8K IOP size, 75% Random, 80% write 8K I/O Alignment.