As promised before, I want to start sharing some of our experiences with our on-going (IBM) VDI reference architecture (RA) work.
I’ve discussed VDI patterns and inhibitors (based on real-world client engagements) in various previous articles so I’ll cut to the chase …
What have we been doing?
We have set up three industry-leading VDI solutions in our labs (US and UK) and are performing architecture verification and LoginVSI performance tests on different sets of IBM hardware, IBM Blades, IBM Rack systems and the recently announced IBM PureSystems (which has great potential to become the ideal platform for VDI – I’ll get into more details in another post).
- Coverage on VMware Blog
- Coverage on Citrix Blog
A key design principle of our Reference Architecture is to address the arguably most common inhibitor to VDI – storage cost. Our approach will radically reduce the storage requirements for your VDI deployment by favouring local SSD storage instead of large-scale shared storage arrays.
We have been working closely with each individual vendor during the project and I’m grateful for their help (since our initial architectural workshop with the vendors last year).
It is also important to point out that the purpose of this document is NOT to compare the vendor solutions AGAINST each other but to demonstrate the suitability of our architectural approach on IBM hardware for each solution.
A key objective for the project is to determine the supported user density for individual workloads in order to create scalable building blocks and sizing models – essentially make sure our approach works, scales and gets you to the best price point.
The three VDI solutions implemented and tested are:
- VMware View 5 (with a 5.1 update to follow) with ESX 5
- Citrix XD 5.6 (ESX 5 host – optional Hyper-V hosts)
- Virtual Bridges Verde 5.5 (KVM based hypervisor with “Storage Optimizer”)
I am frequently asked “will using PCoIP instead of RDP create any overhead?” or “does filling up memory in my server help or harm user density?” …
So what is the impact on user density (and therefore on the all-important metric of cost/user) when:
- Connecting with PCoIP (instead of RDP)
- Connecting with SPICE (instead of RDP)
- Enabling 3D capabilities on View5 desktop pool and image
- Configure an Aero theme for users (in a 3D enabled View 5 pool)?
- Running with decreased memory frequency (when using larger memory configurations)?
- Enabling View 5 Persona Management (compared to Microsoft Roaming Profiles and local profiles on persistent disks)?
So we set out to determine the following values (example View 5 / dual socket Intel Westmere 6-core / 192GB RAM):
You can see that I’ve already listed some of the results as “teasers”. Pay for instance attention to the decrease in supported users when using PCoIP instead of RDP for all users. 20% less users is significant but of course PCoIP provides e.g. advanced graphics and redirection capabilities. Also the results are measured using the default settings (e.g. BTL enabled etc.) so tuning is absolutely possible.
We’lll discuss and share the other results in the following posts, of course all the results (and more) will be also be officially published (IBM Redpaper) as they become available (starting with VMware View), I will post the link(s) to any documents here as well.
Throughout our testing we have performed extensive performance analysis on all aspects (user density, IOPS, latency, network etc) and the findings will be fed into new VDI sizing tools that we’ll make available in due course.
So … a VDI approach that will allow you to reduce your cost per user, allow you to add scalable building blocks as you grow and all bundled with first-hand technical insight and sizing guides based on our testing … sounds interesting? If yes, then feel free to check out the following posts and upcoming publications …
PS A massive credit to the extended IBM team and the VDI vendor teams for their help with this project!
Andreas Groth
