BriForum 2006 Video: Ron Oglesby on the feasibility of running Citrix in a VM - Brian Madden - BrianMadden.com
Brian Madden Logo
Your independent source for desktop virtualization, consumerization, and enterprise mobility management.
Brian Madden's Blog

Past Articles

BriForum 2006 Video: Ron Oglesby on the feasibility of running Citrix in a VM

Written on Aug 15 2006 23,760 views, 27 comments


by Brian Madden

Since this whole Citrix + VMware and VDI thing is a hot topic right now, I thought it would be a good time to look at one of Ron Oglesby’s sessions from BriForum 2006 this past spring. This session, entitled “Citrix, Citrix on VMware, XP Pro on VMware: A performance comparison, and when to use each,” was based on Ron’s real-world testing and some serious number crunching he did earlier this year.

One caveat: This session was before all the dual-core and before VMware’s VI3 product was released, so some of the economic numbers might add up differently.

Either way, it’s a great session and I think provides some timely information to this conversation.

Watch the session now (new window, QuickTime required)

 
 




Our Books


Comments

Allan Harder wrote Citrix on a VM is really THAT WEAK??
on Tue, Aug 15 2006 1:54 PM Link To This Comment
Wow, it's tough to argue against stats like that!
I have put a couple small Prez Serv farms on ESX 2.x and got acceptable performance with an acceptable number of users. However, I don't manage those farms so I have not been able to track their performance.
I would love to see the performance jump if those went direct onto physical.
I did certain tunings to the Citrix servers but not much more than what Ron mentioned early in his presentation.
Anyone been able to compare when jumping up to ESX 3.x? I would love to hear reports of perople using it in production on the ESX 3.x platform ...and not just as one-off servers.
Michael Burke wrote RE: Citrix on a VM is really THAT WEAK??
on Tue, Aug 15 2006 2:19 PM Link To This Comment
Typically, Citrix works well on VMware when the server only services a small number of users.  If the server is being hit heavily, then the vritualization overhead actually hinders performance.

Since ESX 3.0 doesn't really change the underlying architecture of the virtualization layer, performance will probably be the same.
Guest wrote Hmmm...
on Tue, Aug 15 2006 4:15 PM Link To This Comment
I have a Citrix Farm deployed on an ESX 2.5.3 infrastructure...
To the paper tiger comment - need vs. want is the answer (if you need space - server consolidation is an obvious candidate).
 
I have two Racks of space in a datacenter (can't fill them to the hilt because of heat and total amps drawn must be no more than 80% of provisioned power). This basically means ESX servers (Quad Opteron DL585) run ESX with a SAN holding all the VM vmdk's.
 
Each Citrix server can handle 50 users running a database app (big EXE's and lot's of data pulls with each query). Outlook as well (policy and archiving keeps the in memory size of outlook.exe to about 25-25MB per user). Everything is snappy and peppy.
 
I know for a fact that if the VM had the same architecture as the ESX server and was running physically instead of virtual, I would have a screaming farm, but that was not the goal or priority.
 
Performance is more than acceptable and honestly, I believe each server could handle 75 users easy.
 
I have not watched the video yet, but plan to - just wish it were not QuickTime so I would not have to install anything .
 
Brian Madden wrote RE: Hmmm...
on Tue, Aug 15 2006 6:06 PM Link To This Comment
Yeah sorry about that whole QuickTime thing.. In case you're wondering, we had to use QuickTime because we actually have two video streams in this movie--one for the speaker and one for the screen capture. QuickTime is the only format (out of WMV, Real, and Flash) that can keep two streams locked in sync. For the others you have to render them together as a single stream, which means you lose quality on the screen cap side unless you want a several GB movie!

Brian
Ron Oglesby wrote RE: Thanks for posting this Brian
on Tue, Aug 15 2006 10:07 PM Link To This Comment
ORIGINAL: jmsazboy
 Mr. Oglesby stated in the presentation, ...


 
Mr Oglesby... I dont get that much.
 
Anyway, I think VDI Has its place. THe customer I referenced (without name) did it because they (or more correctly someone up stream in the org) had a thing for not wanting to deal with Citrix... Of course Vmware puts them out now as their prime example fo a VDI in use. Though I have never seen a cost comparison of "if we did this SBC instead of VDI, here would be the cost differences". In their case cost obviously didnt matter. In most environments, our bosses want to know where the money is going.
 
Ron
Guest wrote RE: Hmmm...
on Wed, Aug 16 2006 11:49 AM Link To This Comment
Good call on encoding in QT, that was an excellent presentation.  I think Ron made the ultimate point, which is that none of those solutions is the absolute best or worst, but depends very much on whatever a company's requirements are.  It's good to see some actual benchmarks to help back up the scenarios.  When we deploy Citrix on some new blades next month, I'd love to see if I can find his benchmark scripts and post my results in the forums.
Jennifer Henske wrote RE: Thanks for posting this Brian
on Wed, Aug 16 2006 2:23 PM Link To This Comment


Mr Oglesby... I dont get that much.

Ron

 
Yeah, and next it will be "Hey, old man!"
 
Jennifer
Guest wrote RE: Hmmm...
on Wed, Aug 16 2006 3:17 PM Link To This Comment
ESX3 supports 64-bit OS's "experimentally" on 64-bit systems - Apparently "full" support will not be offered until the Intel Vanderpool and AMD Pacifica chips ship. 
 
We regularly steer customers away from running Citrix on ESX, though we are huge ESX proponents for about everything else (inlcuding HA clustering, Exchange, Oracle, SQL, etc.)  Customer decision-making really comes down to do they want to pay for "optimal" or "good enough" - most folks apply the 80/20 rule and go with "good enough" though Citrix is expected to always be snappy - hence keeping it on physical servers.  We're huge fans of the DL585/DL385 combo - and always love the BL25p's for either blade ESX servers (love 4 1000Gb network connections) or Citrix servers.  It's a great setup all around with EMC on the back-end.
Dennis Pennings wrote RE: Hmmm...
on Thu, Aug 17 2006 2:31 AM Link To This Comment
btw, there are some excellent quicktime alternatives. [link=http: for the people who have a hard time installing the big-ass iTunes..
 
just a sidenote..
Quintin Lette wrote RE: Hmmm...
on Thu, Aug 17 2006 5:55 AM Link To This Comment
Great video and content :-) just wish you guys could justify an Aussie BriForum (Unfortunately I can't justify a trip to US/Europe for BriForum - although I keep trying)

it seemed the video ended quite a bit before the encoding did, might be able to trim a bit off the end and save some bandwidth. The video stops at 55:52 (I'd probably cut it from 55:45 at the BriForum logo fullscreen) and the feed goes on to 104:28.
Tord Bergset wrote RE: Citrix on a VM is really THAT WEAK??
on Thu, Aug 17 2006 6:00 AM Link To This Comment
Hi Ron ++.
3.0 performance testing sounds very interesting - is this something that you (or anyone??) can test and give feedback on?
And what about running 64 bit Citrix on 64 bit VMWare - that should be interesting to test...
PS! Unfortunately I don't have a VMWare rack for testing in my basement like you do Ron...
Guest wrote RE: Thanks for posting this Brian
on Thu, Aug 17 2006 4:05 PM Link To This Comment
 Ron / Brian,

Good stuff on the presso it's always been something of interest to me. Quick question, would you expect bolting in Performance Software on top of the XP / Citrix VM to potentially increase the viability of these as a maybe not a more mainstream solution, but definately a more widely considered solution?

I was thinking of AppSense Performance Manager as a possbililty to improve performance / stability / user count on these types of solutions. I know the software cannot run within ESX, but it can run within the guest OS's which allows the OS particularly Citrix to deliver a more consistent quality of service to the users even under load.

What are your thoughts?




Ron Oglesby wrote RE: Thanks for posting this Brian
on Thu, Aug 17 2006 9:40 PM Link To This Comment
I dont think so. If I were to even get away from processor Prioritization within the Guest OS and try to set proc min's for Apps/processes within a VM really I am doing nothing more than setting a MIN that the server/guest OS really has no control over. The guest OS is (if running at 100% cpu that is being allocated) can not get any more CPU. Now I COULD reduce CPU usage in all the VMs thereby allowing for More VMs, but in doing this I slow execution within the VMs and drive time up. Woudl the perfromance be more consitant, probably. But it would me consitantly slower in all VMs.
 
Now If you want to use a perfromance tool to ensure that a run away process doesnt kill the VM and thus hurt (or potentially limit) the amount of CPU cycles to the overall box and thus other VMs, then that is possible. But in the tests we did it wasnt a run away process and no single session was hogging the procs, It really was an even distribution and just overhead on the proc that drove the user count down.
 
NOW if you dont have heavy load on the proc your numbers will be different, but with ESX having its own mechanism for evenly distributing processing power among VMs I dont know that other tools will (within a VM) will help that.
 
Ron
Dennis Pennings wrote RE: Thanks for posting this Brian
on Fri, Aug 18 2006 6:31 AM Link To This Comment
I agree VDI has its place, but only for special circumstances:for those apps who are bad for Citrix, for that company who has to have an offshore programmers team, etc. But i agree with jmsazboy that a comparison versus with Citrix wont hold up. I believe in the comparison before making the decision, and find it hard companies still make a dicision without all the facts of all the options on the table. Sure, that company DID choose for this solution, but i cant helping they didnt make a balanced dicision, and they decided mostly based on politics, as it always is in big companies. I wonder if u made 2 solutions (Citrix vs VDI), with all the facts, including $$ and focussing on a solution on their problem, if YOU would have made that decision. I bet u would have created a solution with Citrix AND VDI..
Dennis Pennings wrote there goes our solution to our prob..
on Fri, Aug 18 2006 6:36 AM Link To This Comment
In my world, the biggest limit is the 2Gb limit of terminal services. 64bit is too early for us and our customers, so we depend on 32bits. Our customers use a 3rd party app for their daily business, which consumes about 10-100mb per user, the biggest Outlookfanssessions use a similar amount, and we have some Excel/access (fanatic) users. Which leaves the amount of users per server on about 20-30 users till the memory runs out (with some memory tuning). 30 users is acceptable in my world, but 20 is not if it relies on physical hardware. And we see a steadily rise in usage per server as the newer versions of the software (and office) are being used, so in a year or so, we might only be able to use 15users or so. This becomes too expensive to buy and manage. This is where VMware ESX seemed to be my solution, but according to this presentation, its a NoGo. O well, at least you saved us the testingproject. ;-)
Dennis Pennings wrote RE: Thanks for posting this Brian
on Fri, Aug 18 2006 6:49 AM Link To This Comment
Mayby i missed it (i remember a statement about the dualcore), but can u answer me this; Can u elaborate on why Citrix within VMware is slower?
 
And do you expect it to scale linear? I mean, if i would invest in more procs (mayby in a 4proc-dualcore ESX) per VM, do u expect it to have the double of users? In this case, ESX ifor Citrix is still an option, just invest in more procs, right?
 
What has VMware to say about your experiment?
Jeff Pitsch wrote RE: there goes our solution to our prob..
on Fri, Aug 18 2006 7:41 AM Link To This Comment
Have you actually perfmon'd this and confirmed it or are you simply estimating th ememory based on task manager?  I ask because what you see in taskmanager is not the correct when it comes to memory?  Windows will actually share the memory space of the executables.  For instance, the first person to run Outlook loads it into memory, every other user that starts outlook then points to that executable, they do not get a separate outlook.exe in memory.  So if you have 15 users start outlook, the memory consumption is 1x25meg, not 15x25meg.  Now if a write happens to that executable, then a copy is made of it so then you can get multiple instances of a file in memory but my whole point is,if you haven't actually stressed tested with users to see how many users you get, then you may be able actually get more on the systems.
 
Hopefully that made sense....
Dennis Pennings wrote RE: there goes our solution to our prob..
on Fri, Aug 18 2006 9:29 AM Link To This Comment
yeah we used taskman to see usage and then did a thourough investigation with perfmon.. it was actually before we found out about the 2G limit and tested this on a live farm where there were 250users with about 40users per server (which crippled the servers).. We saw the 2g limit reached at about 25users (which was about 3years ago!) and then the page file usage went up, even though there was 2G available.. then brian explained me about the 2G limit.. But my point is, 25users with that app + office/outlook, 3 years ago was enough to reach the server limit.
 
Btw, we found out that the amount of memory used in Outlook is related on how much stuff there is inside (public folders/tasks/agenda/etc), so i figure (correct me if im wrong) that that goes beyond the amount used by the executable..
Ron Oglesby wrote RE: Thanks for posting this Brian
on Mon, Aug 21 2006 10:04 AM Link To This Comment
ORIGINAL: Dennis@7summitsict

Mayby i missed it (i remember a statement about the dualcore), but can u answer me this; Can u elaborate on why Citrix within VMware is slower?

And do you expect it to scale linear? I mean, if i would invest in more procs (mayby in a 4proc-dualcore ESX) per VM, do u expect it to have the double of users? In this case, ESX ifor Citrix is still an option, just invest in more procs, right?

What has VMware to say about your experiment?

 
Dual cores when you have a large number of ring 0 calls or a large amount of utilization could give you more sessions per phsyical box. Now what that means (to get double or close) would really mean more Citrix VMs on the same box spreading laod across more cores. In my test with proc being the bottle neck I would look at a 60-70% increase if I added more VMs.   If proc WASNT the bottleneck I would see much of an increase in all.
 
Ron
Ron Oglesby wrote RE: there goes our solution to our prob..
on Mon, Aug 21 2006 9:17 PM Link To This Comment
ORIGINAL: Dennis@7summitsict

yeah we used taskman to see usage and then did a thourough investigation with perfmon.. it was actually before we found out about the 2G limit and tested this on a live farm where there were 250users with about 40users per server (which crippled the servers).. We saw the 2g limit reached at about 25users (which was about 3years ago!) and then the page file usage went up, even though there was 2G available.. then brian explained me about the 2G limit.. But my point is, 25users with that app + office/outlook, 3 years ago was enough to reach the server limit.

Btw, we found out that the amount of memory used in Outlook is related on how much stuff there is inside (public folders/tasks/agenda/etc), so i figure (correct me if im wrong) that that goes beyond the amount used by the executable..

 
It may still be the solution to your problem. Remember I used a laod script that loaded both processor and mem. If you are hitting the memory limits in windows AND your processor utilization is low you still may be in for a solution. I mean imagine if you get 30 users on a single proc with the 2 gb limit then trying to virtualize might be the thing. Sure you may drop a few users per VM but you could increase the numebr of user PER BOX.
 
This is a classic "When to use Citrix on VMware" Scenario. Funny thing is this doesnt happen as much as people talk about it.
 
Ron
Guest wrote RE: there goes our solution to our prob..
on Tue, Sep 5 2006 9:08 PM Link To This Comment
Ron:
Are you working on VMWare Infrastructure 3 book?
When will it be out?
Ron Oglesby wrote RE: there goes our solution to our prob..
on Wed, Oct 18 2006 4:09 PM Link To This Comment
ORIGINAL: Guest

Ron:
Are you working on VMWare Infrastructure 3 book?
When will it be out?

 
yes I am. As soon as Scott, Mike and I can get it OUT! mid 2007.
Joe Zalinsky wrote RE: there goes our solution to our prob..
on Thu, Mar 1 2007 1:38 PM Link To This Comment
ORIGINAL: Jeff Pitsch

Have you actually perfmon'd this and confirmed it or are you simply estimating th ememory based on task manager?  I ask because what you see in taskmanager is not the correct when it comes to memory?  Windows will actually share the memory space of the executables.  For instance, the first person to run Outlook loads it into memory, every other user that starts outlook then points to that executable, they do not get a separate outlook.exe in memory.  So if you have 15 users start outlook, the memory consumption is 1x25meg, not 15x25meg.  Now if a write happens to that executable, then a copy is made of it so then you can get multiple instances of a file in memory but my whole point is,if you haven't actually stressed tested with users to see how many users you get, then you may be able actually get more on the systems.

Hopefully that made sense....

 
It all makes sense... if only it were true.
 
I am currently looking at a unit where I have 5 users with outlook loaded... and as I count them there are 5 instances running all using real memory and all using additional virtual memory.
 
JZ
Guest wrote broken link
on Tue, May 29 2007 5:18 AM Link To This Comment
Can someone fix the link!
Guest wrote Re: broken link
on Sat, Sep 13 2008 3:34 PM Link To This Comment

I agree please fix this link

(Note: You must be logged in to post a comment.)

If you log in and nothing happens, delete your cookies from BrianMadden.com and try again. Sorry about that, but we had to make a one-time change to the cookie path when we migrated web servers.