IOPS requirements

Hi Robb,

Can you post the link to the whitepaper here?

Shawn, I think I have been hearing/reading that memory has been really well addressed in the UCS model however I have not been seeing much about improving the IOP bottleneck as it relates to VDI and shared storage.  Like Michel, I would like to see some specs.  As you are well aware we have clients who are very interested in solutions that can address this issue.  Since so few technologies seem to be able to address the IO limitation ...the argument for local storage is the scalable one-today:)

I liked this chart where he compares OS with various memory configs and hardware and the resulting max users.

[1] The number of cores specified on the Xenon E5530 represents the physical number of cores on the machine and not the number of cores reported by the Operating System.

[2] This is based on the stated support limits for Hyper-V. Specifically this takes into account that the maximum supported ratio of Virtual Processors to Logical Processors is 8:1.

[3] This is the maximum number of users the system was able to support without a significant degradation to user experience. This limit was not always determined based on a degradation of user experience; in some cases this was simple the point where no more virtual machines could be started due to lack of memory.

I believe it was from Michael Kleef and was back in March... here is the link to the whitepaper.  He does a nice job of effectively comparing xp vs win7 as guest os for vdi.  http://www.microsoft.com/downloads/en/details.aspx?displaylang=en&FamilyID=bd24503e-b8b7-4b5b-9a86-af03ac5332c8  .  Jon I dont think in this article that Michael has tweaked the OS's.  

Yes, I think I read that paper before. It is a very good one. Windows 7, after boot actually can do better from a IO perspective than Windows XP does. If you can deal with the windows 7 read I/O storm at bootup, Windows 7 would actually perform better. Watch ProjectVRC the next couple of days for more cool information on this.

Hello Robb!  IOPs are most def. an issue.  I think when you go to extreme measures to physically consolidate VDI/TS environments, you also have to do this with storage.  At the end of the day, it's all about how fast can you get that data.  In order to scale to extreme numbers, you likely have to look at SSD/NAND type storage architecture.  We've looked at PCIe options and SAN attached NAND options.  Both give extreme (100,000+ IOPS) numbers that allow this kind of scale.  Here is the link I was referring to.  Again, it's not the gospel...it's just a white paper. http://www.cisco.com/en/US/docs/solutions/Enterprise/Data_Center/App_Networking/vdiucswp.html

Sorry for the delay...I removed the notifications for this topic.

I don't know that I completely agree.  "It depends"  If you have to purchase enough spindles to gain 100,000 IOPs... what does that cost you?  We obviously aren't talking about small projects.  Where talking about spanning multiple racks of fiber spindles in comparison..  That adds up.

Smaller storage capacity though - for the money.

Wouldn't the memory requirement be higher though?

Local storage is compelling with the proper design and strategy, but that's a religious topic.  My question is why one would spend all that money for a dozen spindles when the same cash can purchase 100 times the IO capability with SSD/NAND technology.  Smaller footprint, less heat, less energy, etc etc..

What are your thoughts Robb 

Likely 175-225 depending on who you ask.  185 is the number I've always used for "calculations".

Yes, very much so. One could argue that there could be perceived to be a sweetspot for a economical server that has the memory and the IOPS to host that many Windows7 desktops. WindowsXP would hit the IOPS bottleneck earlier or at the same time with a lot of RAM left.

I knew I forgot to do something....

"theoretical IOPS capacity is for the IBM 146GB 15K 6Gbps SAS 2.5" SFF Slim-HS HDD"

Is 160 - 180 per drive.

While I will agree.. that IOPS number is going to depend on if we are talking about a single drive, Raid1(10), Raid 5, Etc... Depending on Raid level, your reads and writes are going to be handled differently thus changing the "formula" needed to calculate IOPS.