Is there a "white paper" on financial savings to show how these percentages are figured out?
Ron Oglesby: We have a model we use for Cost avoidance on initial Capital investments there is a whitepaper on Rapidapp’s site ( about the assumptions you have to calculate it is base don upgrading hardware for X number of NT 4.0 servers (project based).
In addition the IDC reports show an average server uses about $23 in power per month and $40 in cooling per month. Plus they estimate that hardware support (not OS or software) is about 1300 per year per server. Add that with ongoing network upgrade costs (additional switches mean more IOSs to support etc.) You get about $2300 per year per server (NOT PER OS just hardware).
Why does everyone assume a VM doesn't cost anything to support? I can understand capital costs being 1/2 - 2/3 less, but labor costs are not very different than a physical (patching, security, etc.)
Ron Oglesby: No one assumes that. Patching and what not is the same. As a matter of fact VMs are so much cheaper in capital up front cost you often end up with MORE OS’s. but as stated in the previous question, the IDC says the industry average for Hardware support, Power and HVAC is about 2300 US per server. No Software or Personnel costs built into that. the ongoing support price (even though techies don’t often get it) is lower.
I was told that Microsoft does not support VMware, so how can I implement VMware..
Ron Oglesby: MS gives best as can support. Meaning if you call up with a performance problem on a VMware VM they may tell you it’s a VMware problem after doing some initial troubleshooting. if it is something like a replication issue that is purely software, they support it. Also if this is a major concern you can mitigate the risk by buying your VMware licenses from HP or IBM and gaining both VMware and MS support from the vendor. they offer it. Or create a V2P migration (virtual to physical) process for moving a VM to a physical box for that type of troubleshooting.
Curious, does ESX running on Opteron 64 support Windows x64 as a guest machine? I know that VS does not.
Not yet but it’s on the road map
How would you monitor performance of a thin client to Citrix servers, if there is only a chip with ica on it?
We monitor the session latency from agent installed on the Citrix server
How large will a sql 2000 server grow in a month by monitoring 1000 citrix users?
This would depend on the number of Citrix servers and monitored application.
What are you using to measure latency? SRTT?? ICMP??
We measure latency from the ICA session and back-end response by monitoring requests being sent to back end servers
Can you give us an example for improved performance with VM? A Software Product, what ever?
Ron Oglesby: There are two ways performance is IMPROVED. The first is that the server is being migrated to newer hardware, faster processors etc. that in itself will sometimes improve performance just as it would if you moved to a new physical server. Another time you can see an improvement is if you are running into OS limitations, virtual address limits or file system cache limits etc. In this case you are increasing the number of OSs on a singe piece of hardware to better use it and get more transactions out of the box. This of course is not a “real” performance increase. A VM is still using hardware and it all comes back to the hardware, there is no magic to make things perform better other than adding newer hardware or more OS’s to the same box.
Can you talk about the relationship of Dual Procs on VM's with the hyperthreading technologies in server platforms (HP DL 360/380)
Ron Oglesby:Hyperthreading in VMware does show a 20-30% boost in perf on the processor side. The Vmkernel attempts to take this into consideration by not allowing a dual proc VM to be run on the same Physical processor. The VM itself is not aware of the hyper threading all it sees are two virtual processors that are being presented by the VM kernel
Any results or experiences with Citrix/AppSense in VMWare environments?
Ron Oglesby: Works just like the others. Same types of results, though everyone has their favorite performance tools
What was that tool? Thumbtak or Thumbtech?
Performance Guard from PremiTech
Ron Oglesby: or is he talking about Aurema’s armtech?
You talk about users/physical server recommendations. I was just at an organization that had 120/server. Now they did start to have load issues. Qfarm stats were at @8500-9000. We added a few more servers and brought it down to @5000. Now, their apps are all home grown (VB, Power Builder, Java) and don't require Windows based installations. What affect do these typpes of apps have on user count per server.
Ron Oglesby: exactly right. the application in Citrix are really what drives the user per server or user per procs, The only time the OS affects it is when you run into logical limitations of the OS design. Saying an app is written in VB or Java etc doesn’t really tell you how the user count will perform. Now as a rule large apps written in VB tend to be more sloppily written and therefore have more overhead than something like… Office, which is actually pretty tight. But the app its self, its design and the type of hardware is really going to be the deciding factor.
How do you judge a Piii 900mhz vs a zeon p4 3.5 clock speed. Is a cpu cyce still a cpu cycle
Ron Oglesby: Not sure what this means, but I do agree. a cpu tick is a tick, the speeds only say I can have so many ticks a second. In the graphs in the session I was showing VMs configured two different ways on the same hardware. So the comparison was apples to apples. If I was comparing utilization from a 1GHz phys box to a VM on a 3.0 xeon then it would be apples to oranges since there is three x the processing power. Not sure this is relevant other than the fact that moving an OS that was on 1.0 ghz to a 3ghz machine may INCREASE performance. see previous questions.
Is there any value to perfmon counters measured inside ESX ?
Ron Oglesby: Sure, some counters are still valid. Solid number counters like Processor queue length, disk queue length, etc. its counters that are based on the system clock that are effected by time drift. Things like % of disk time, % of proc utilization, Context switches/sec etc.
One of the slides refered to Vanderpool what is that product do for conterxrt switches
Ron Oglesby: Vanderpool (intel’s future) will have an effect on the current processor “thunking” I described in the presentation. Right now when a VM needs to run a kernel mode process (like establishing a TCP connection) it needs to be done at Ring level 0 in the processor. This requires that ESX allow the VM time at that level then execute the instruction, then reclaim the position in level 0. Its akin to thunking, also called a context switch. Vanderpool will move those types of functions out of ring 0 allowing the VM to execute those without interfering with the VMkernel. huge increase in performance for servers with large numbers of kernel instructions (visualize an internet facing SMTP server with thousands and thousands of messages a day)
Does the cost of users/CPU tend to go down if running on Citrix in VMware on an 8-way?
Ron Oglesby: Not really. unless you are already hitting the OS limitation on your Citrix server the cost per user with citrix generally goes UP with Citrix VMs. Assuming you are already pushing the hardware and move the servers to a VM you now need to increase the number of OS’s you are running (back to the single proc VM example for performance reasons). So lets use a dual as an example. If it was physical you had one OS license. Now if you have one VM per proc (because you are using the hardware) you now have 2 OS licenses one for each VM, or in an eightway, 8, one for each VM for each processor. Add that to the fact that cost for hardware is not linear as you move from dual to quad to eight-way, and the esx licenses and you have a MORE expensive solution per user.
Of course if you are NOT running into hardware limitations and instead have an RSL limit or a limit of system PTEs etc… Then you can scale up. Maybe you have 50 users per dual but are hitting the PTE limits for some reason (not likely, but lets run with it). You have tested and found you can create 3 VMs on that same dual and get 35 users per VM. That is 50 Users physical vs. 105 on VMs on the same hardware. You have effectively doubled your capacity. In the cases of OS limitations (now becoming fewer and fewer with 2003 Srvr) you can increase users per server and decrease costs.
Don’t get me wrong there are other reasons to go VM, Maybe you need the recovery options for DR, or faster provisioning and what not. But with a primary silo Citrix server (or any server) that is truly using its hardware will require that you pay a price for those benefits. With that said., The last three analysis I did of companies with more than 1000 servers.. 84% of the servers never reach 10% proc utilization during working hours. Something to think about.