Poor VMWare View 4.6 Performance

[TrueBrit]

We have deployed a small VMWare View solution for Admin Staff, Accounting and CAD / Drafting users. We currently have 21 VM's running and received daily complaints about performance, some days better than others but overall well below the performance level of an appropriately configured Desktop. We are either needing to end the production use of this system and return back to standard desktops or add an additional VMWare view server and soldier on, hoping we can find the cause of the low performance.

Here are the spec's.

Server Running VM's:

Dell PE R510, 2 x 6 core Xeon 5670 CPU's running at 2.933ghz, 64gb of Memory, 1.81 TB Raid 5 Storage using 6 SATA 6g 7200RPM drives, 4 x 1gbit Network cards

VM Config:

Windows 7 32bit, 2 CPU's and 3gb memory per system, we have 3 Linked Clone pools for Accounting, Admin and CAD specific software installs.

I just cannot see were the loss in performance is. Sometimes users sessions hang for a few seconds when they select various functions in office or AutoCad, then the system becomes responsive again. Other times the screen quality degrades a little before clearing up. For some office programs take an excessive amount of time to load and become available. Sometimes login or general file / network access is slow, folder contents lists taking prolonged periods to refresh or display contents. Sometimes functions such as printing and Plotting take excessive amounts of time before their respective dialog boxes appear.

I have attached screen captures of our performance stats for review. Any assistance would be most welcome.

[AGratefulDad]

Hi there,

I hope to give you some insight on perhaps some of your issues.

1. RAID 5? This is not a suggested RAID set for VDI unless you have a ton of spindles (20+ minimum)

2. SATA? You are only getting 80 IOPS out of a SATA drive on the upper end vs. 170+ on a SAS 15K Drive

3. Network - how do you have your network interfaces broken out? 2 Virtual Machines / 1 Managment / 1 vMotion?

Remember, VDI is about 70%(80%) write and 30%(20%) read in terms of your I/O calculation. How many spindles do you have in your RAID set?

More information would be great to help you out even further.

I speak from experience, we made some pretty serious errors when rolling out our 50 seat View 4.6. I asked a lot fo questions way to late, but I listened, followed the recommendations and today we have a successful VDI infrastructure that is constantly improving.

Also - looking at your disk latency, you are waaaay over the "acceptable" standard which would tell me your session are starved for disk I/O resulting in freezing and serious performance problems.

Cheers!

TheVirtualGuy

[Linjo]

I agree, most likely storage is your problem. Add more or faster storage,

// Linjo

[TrueBrit]

Thank you for your reply!

Great questions. I hope I am clear in my answers.

1. On the RAID config, yes I have started to think that is an issue. We have 6 disks in our raid 5 Array

2. Is SAS acceptable then or should we go back to SCSI?

3. We have 1 nic assigned for all the CAD workstations, 1 for Accounting and Admin and 1 for Management

I don’t think we asked enough questions when we rolled this out either. In fact we were forced to go to production much faster than we should have due to rapid hiring and expansion. Now we are trying to determine if we should take the users back to traditional desktops. We have approval for a new implementation on a new server but I am hesitant because I don’t want to make the same mistakes as before. Can you recommend specific hardware configuration or give me insight as to what worked for your environment?

Thank you again for coming to our aid.

M

Matt Steel

CTO, B&T Engineering

[AGratefulDad · Accepted solution]

Hi Matt,

Thanks for providing more information.

On a typical hardware RAID 5 with 6 spindles, your IOP count on a RAID 5 is around 130 IOPs based on a 80% Write / 20% Read ratio.

You are using 21 desktops over that 130 IOPs, so you are essentially allowing for 6 IOPs per desktop without any room for overhead.

Now, if you moved that to 15K SAS with the same amount of spindles, you would end up with a result of 304 IOPs based on the same ratio.

21 desktops spread over 304 is 14 IOPS per desktop, but again, no overhead.

Essentially, just by moving to 15K SAS, you have more than doubled your IO, but your RAID set is not optimal.

If you went Hardware RAID 10 with the same spindle count, your IOPs would increase to 555 based on the same ratios.

21 desktops over 555 IOPs is 26 IOPs per desktop which is getting closer to the 30-40 recommended for VDI.

If you can afford to go 8 x 600GB 15K on your R510's, you'll be pleasently surprised with the performance - it would net you an IOP count of 740

21 desktops over 740 IOPs is 35 IOPs per desktop which is more inline with the recommendation of VDI.

I scoped mine out at 40 IOPs per desktop plus overhead and we have no issues on our end here.

Don't give up, with a little more work and planning, you could end up with a great platform - we use the R510's here as well and love them.

I hope that helps you out, this is a great place to ask questions and get answers.

I am also on Twitter if you want to ask any further questions as well, more than happy to help you out!

Cheers!

[jsavoor]

Hi Matt

I just read your post and I think that you should look LiquidwareLabs and their capacity planning and optimization tool Stratusphere.

See if this solves the problem for you. They also have a superb profile migration product called ProfileUnity.

Jay

[TrueBrit]

Jason,

Thanks for the great feedback!! Everything you said makes perfect sense and I am definitely going to configure the new server differently. If I may I would like to ask a follow up question. If we were to reconfigure the current View VDI server and change its 6 disk RAID 5 to a RAID 0+1 (mirror of stripes) would the speed gain be sufficient to alleviate the immediate suffering until the new server is in and online? Would it be worth the work?

Thanks again for your help it has been much appreciated.

M

Matt Steel

CTO, B&T Engineering

[AGratefulDad]

Hi Matt,

No problem, happy to help.

As for changing the configuration of the current RAID set, that's a tough call, but in my opinion, I would say no - it would not be worth the effort as you will only improve your overall VM performance from 6 IOPs currently to 10/11 IOPs per desktop.

I think if I were in your position, I would focus strictly on getting new disk online as soon as you can. It will be the most inexpensive way to gain signifigant performance. You have a got plenty of CPU and plenty of RAM, it's your disk that is holding you up.

I hope that answers your question.

TheVirtualGuy

[TrueBrit]

Jason,

Thanks again for the insight. I did some checking, we cannot upgrade disks and reconfigure the RAID on our current server due the nature of the backplane we have. So we are looking to either purchase and iSCSI SAN solution and connect it as Direct Attached iSCSI to our current VDI server or replace the entire server for similar cost getting the correct array and doubling the memory. Have you had any experience using iSCSI SAN with VDI? Is there a IOPS performance loss with this solution?

Thanks again,

M

[AGratefulDad]

Hi Matt,

The R510's are a great product, you just got stuck with the wrong back plane.

As for iSCSI SAN's - great option, but what are the specs of the unit you are considering? MD3200 Series? Equalogic? EMC? etc...

Knowing that would allow me to give you some further insight, but yes, you are definatley on the right track by moving to an iSCSI based SAN!

Cheers!

[TrueBrit]

Jason,

You are right on the money. We are in fact looking at a Dell MD3600i configured with 6 x 600gb 15k SAS drives (Not optimum I know, on the number of Disks) configured for RAID 10. This unit comes in at $10k or so. For almost the same price we can get another PE R510 configured with 12 x 300GB 15k SAS drives, double the memory and a little jump in processor power over what we currently have now.

I realize we have much greater expansion options with a MD3600i but we would still have to increase our current servers 64gb of memory to keep up with planned growth. This would add additional $$ to the cost.

Going with the additional R510 would allow us to run at least double our current number of VDI's to about 45 or so if I am correct on IOPS.

But I am open to ideas.

Thanks again,

M

[AGratefulDad]

Hi Matt,

An iSCSI SAN with 6 x 600GB is going to net you around 524 IOPs or roughly 25 based on your current 21 desktops - reasonable based on current desktop load.

A R510 with 12 x 300GB is going to net you 1048 IOPs or roughly 50 IOPs based on 21 or 24 IOPs for 45 desktops as you have suggested.

As VDI IOPs are random and never consistent, you would probably have better success with a single R510 maxed with all 15K drives - but you give up the "shared" storage scenario which in turn limits your DRS / HA and FT capability.

You need to decide what you can and cannot live without - if clustering and other HA / DRS and FT features are not important to your infrastructure, then the R510 would net you higher performance. If you have plans on enabling cluster features such as those listed above, then I would suggest you move to the MD3600 to take advantage of them and add disk as your budget allows.

Again, it comes down to what your "whole plan" looks like, from there, you can start eliminating options.

Let me know if I can be any further assistance.

Cheers!