Lance Boley: Good day everyone, we are just about ready to get started ... just waiting on a few more to join.
Dell-Yinglong: Hello, Yinglong here from DVS
Lance Boley: Welcome ...
Lance Boley: Okay lets get started ...
Lance Boley: Today we are going to have some experts on to talk about Dell's Desktop Virtualization and what-if's around Office 2013 and Boot on demand.
Lance Boley: John, maybe you could introduce yourself and tell us about the team today.
Dell-John Kelly: Thanks Lance
Dell-John Kelly: I work in the Cloud Client Computing Engineering organization
Dell-John Kelly: we are responsible for delivering aoo of our end-to-end desktop virtualization solutions
Dell-John Kelly: My main remit is around performance analaysis and characterization (PAAC)
Dell-John Kelly: Recently we've looked at 2 interesting topics
Dell-John Kelly: Number 1 - what are the density implications for moving to Office 2013 (in terms of number of virtual desktops per compute host)?
Dell-John Kelly: Number 2 - What are the performance implications of booting virtual desktops on demand i.e. not having all desktops on a VDI compute host pre-booted?
Dell-John Kelly: Not surprisingly I will concentrate on number 1 first : -)
Dell-John Kelly: We saw some interesting results
Dell-John Kelly: Summary is that Office 2013 has a big impact on density compared to Office 2007 and Office 2010
Dell-John Kelly: Office 2010 and 2007 are practically identical to each other
Dell-John Kelly: Let me just bring up the exact numbers here to give you more specifics....
Dell-John Kelly: With Office 2007 or 2010, we can get 116 knowledge users
Dell-John Kelly: But guess what - for Office 2013, it's 95.
Dell-John Kelly: That's a per host number (R720)
Dell-John Kelly: That's approximately an 18% hit on density!!
Dell-John Kelly: TWhen I say knowledge worker I mean a Login VSI medium user
Dell-John Kelly: Which is pretty much the industry standard knowledge worker
Dell-John Kelly: But for Login VSI light (task worker) - something really weird happens....
Dell-John Kelly: We got a denisty of 95 also for Office 2013
Dell-John Kelly: Comarped to 145 on Office 2007 / 2010
Dell-John Kelly: ^ Compared
Dell-John Kelly: So we re-ran our tests and scratched our heads
Dell-John Kelly: But still got the same, our results were sound
Dell-John Kelly: It seems to us that the way Login VSI have put together their workloads, both Light and Medium actually have the same amount of Microsoft Word activity
Dell-John Kelly: The difference is the extra activity coming from other acitivites (e.g. flash video, Excel, Powerpoint etc)
Dell-John Kelly: But Word is driving the "high watermark"
Dell-John Kelly: But I must say it still concerns us that on the latest version of Office, Login VSI 3 gave us this result
Dell-John Kelly: But it seems like help is at hand as Login VSI have re-constituted their workloads in Login VSI4
Dell-John Kelly: We haven't done any testing on Office 2013 / Login VSI4 yet ... but am eager to get around to it soon
Dell-John Kelly: But the main takeaway is - migrating to Office 2013 has serious density implications in terms of per server density
Dell-John Kelly: anything from 20-30%
Dell-John Kelly: So you need to think hard about whether you need the functionality
Dell-John Kelly: I want to give my fingers a rest
Dell-John Kelly: Any questions or comments?
Dell-John Kelly: Going once...
Lance Boley: Tom ... nothing from you today?
Lance Boley: or Greg?
hypervfan: One thing is doing testing with software like that, but have you seen more memory/cpu usage when testing with just a few users?
Dell-John Kelly: memory and cpu scales quite linearly is our experience
Dell-John Kelly: But of course as the users log in, the desktops are all pre-booted and sitting idle, so that will take up some cycles to start with obviously
hypervfan: And Office 2013 goes up faster than 2010?
Dell-John Kelly: in terms of cpu usage, it sure does
hypervfan: Is all your testing done with software, or is some manul testing done as well?
Dell-John Kelly: we normally login all users using login vsi but we also log in one user manually and that manual user executes sample tasks
Dell-John Kelly: because a human being is the ultimate judge of end-user experience...
Dell-John Kelly: It's important to say that we also use the Liquidware Labs Stratusphere UX tool to monitor end-user experience
Dell-John Kelly: that looks at things like login time, application load time etc
Dell-John Kelly: so tools-measured end-user experience, human-measured end-user experience and resource utilization monitoring (CPU, RAM, IOPS and network) gives us a good holistic perspective onthe environment
Dell-John Kelly: any other questions? Have I answered all of the questions to date comprehensively?
Rasmus Haslund: Hi all
Rasmus Haslund: Sorry for the late arrival
hypervfan: How is IO on Office 2013 compared to 2010?
Dell-John Kelly: We saw it was a little higher
Lance Boley: Hey Rasmus ... glad you could join us
hypervfan: Hi Rasmus
Rasmus Haslund: What caused the 18% drop in density? Memory usage? CPU usage? IOPS?
Dell-John Kelly: The drop in density was due to increased CPU
Dell-John Kelly: The IOPS increase was of the order of 10-20%
Rasmus Haslund: Was the increased CPU usage across *all* Office applications?
Dell-John Kelly: IT was mainly Word
Dell-John Kelly: From what we could see, small delta in the others - but mainly Word
Rasmus Haslund: What about memory usage?
Dell-John Kelly: We concentrated on CPU because that is the gating factor that drove density; I don't have the memory information to hand
Dell-John Kelly: If I had to make an educated guess, I would say probably a little higher
Rasmus Haslund: For me thats interesting you say that, I have made VDI deployments using Citrix VDI-in-a-box and usually CPU was not 1st or 2nd constraint
Dell-John Kelly: Our experience has usually been that CPU is gating factor in our testing
Dell-John Kelly: But that is with our standard R720 compute host config of 192GB
Dell-John Kelly: So no overcommitment for taks or knowledge workers
Dell-John Kelly: ^task
Dell-John Kelly: Of course if we went down to 96GB that would probably be different
Rasmus Haslund: Did you see if Office 2013 was better at using SMP than Office 2010/2007?
GC: are there any plans to test this in XenApp or VMware ThinApp scenarios. Knowing they are not quite the same kind of products but curious.
Dell-John Kelly: We didn't drill down into advanced CPU features such as SMP; but I do know that MSFT are making their Office apps more multi-core /multi-proc aware, so I would guess Office 2013 may make better use of SMP
Rasmus Haslund: I dont have more questions on subject #1 for now, looking forward to hearing about #2 : )
Dell-John Kelly: Currently no plans but (not sure if you're internal or external to Dell) - if you feed the request in through the appropriate channels, it may be possible
hypervfan: Looking forward to topic 2 as well : )
Rasmus Haslund: We could ask Dennis and Lance to feed it ;)
Dell-John Kelly: One other comment - Citrix did similar testing on XenDesktop and got similar results
Rasmus Haslund: Did you test this on both Windows 7 and 8? Similar results?
Dell-John Kelly: Windows 7 only
Dell-John Kelly: One other comment is that it was a one vcpu virtual desktop - industry is moving to 2 vcpus for higher-end virtual desktops and Dell will of course lead that move
Dell-John Kelly: Including similar testing to this
Rasmus Haslund: I would be really interested to know the results of this test, but adjusted to 2 vCPU
Rasmus Haslund: Any chance for a follow-up report in the Rockstar forum on this?
Dell-John Kelly: I can't commit at this time but I will take your request onboard
Rasmus Haslund: Great : ) Time is running fast I see... subject #2 hopefully soon : )
Dell-John Kelly: Login VSI's new workloads in Login VSI4 are targeted at 2 vcpus for knowledge worker - so you will most likely see a lot of the results you are looking for very soon
Dell-John Kelly: Agreed
Dell-John Kelly: So - boot on demand
Dell-John Kelly: Scripting and po wer management technologies allow all desktops to be 100% pre-booted on a "just-in-time" basis
Dell-John Kelly: But still some customers want to know how their environment will work in a boot on demand environment
Dell-John Kelly: so we tested our local tier-1 solution (tier1 storage - where the virtual desktops reside)
Dell-John Kelly: with a number of different scenarios
Dell-John Kelly: 100% pre-booted (i.e. fully pre-booted); 20% pre-booted and 5% pre-booted
Dell-John Kelly: 5% is quite aggressive
Dell-John Kelly: And logged all users in over the space of one hour
Dell-John Kelly: To replicate a normal morning login storm
Dell-John Kelly: And we were actually very happy with our results
Dell-John Kelly: Both from the perspective of lstorage atency and Stratusphere UX end-user experience
Dell-John Kelly: ^ storage
Dell-John Kelly: latency was very comparable across all 3 scenarios
Dell-John Kelly: 6 milliseconds for 100%, 7 milliseconds for the other 2
Dell-John Kelly: which is well below the 20ms latency that's commonly used in the industry as rule of thumb for acceptable latency
Dell-John Kelly: Stratusphere UZ also showed consistently good end-user experience
Dell-John Kelly: ^ UX
Dell-John Kelly: IOPS per user not surprisingly increased
Dell-John Kelly: For 100% it was 10 IOPS
Dell-John Kelly: For 20% it was 12 IOPS
Dell-John Kelly: and for 5% it was 13 IOPS
Dell-John Kelly: As expected because the extra booting is an IOPS intensive activity
Dell-John Kelly: Also not surprisingly - we saw a change in the read / write percentages for the IO
Dell-John Kelly: 22/78 for 100% pre-boot
Rasmus Haslund: This might be a stupid question, but I assume this is for non-persistant desktops?
Dell-John Kelly: 37/63 for 5% pre-boot
Dell-John Kelly: First it's not a stupid question, I've never seen a stupid question"
Dell-John Kelly: Second, it was non-persistent, that's correct
Dell-John Kelly: The change in read-write percentage is also as expected - because booting is a read IO intensive activity
Dell-John Kelly: We have 4 minutes left
Dell-John Kelly: So questions please
hypervfan: read-write percentage is for that hour of testing?
Dell-John Kelly: correct
Dell-John Kelly: we totalled them across the testing and got the ratio from that
Dell-John Kelly: we log all desktops in over an hour, then run them at steady-state for 15 minutes, then log them off
Dell-John Kelly: note that this is local tier 1
Rasmus Haslund: Maybe a definition of local tier 1?
Dell-John Kelly: tier 1 is local to the R720
Dell-John Kelly: i.e. sitting behind a PERC
Dell-John Kelly: not on a SAN
Dell-John Kelly: RAID 10
Dell-John Kelly: one other point is that the increased IOPS per user will obviously have an impact on the n umber of users that will "fit" on a SAN (although we didn't test on SAN)
Rasmus Haslund: What impact do you expect features such as vFlash to have on these designs?
Dell-John Kelly: any flash type functionality will obviously increase capability in relation to number of desktops that can be booted with good performance in a particular time window
Lance Boley: We are at the top of the hour and just about ready to wrap it up for today.
Lance Boley: So are there any last questions ?
GC: Thanks John. Great job explaining this body of work!
hypervfan: Great info
Rasmus Haslund: Great chat tonight
hypervfan: Interesting to see the 2 vCPU tests when they are coming : )
Rasmus Haslund: Hoping to see John back again another time!
Dell-John Kelly: hope you guys found it useful and I will try to incorporate your suggestions into our future work
Lance Boley: Alright, thanks everyone for joining today, talk to you next time.
Dell-John Kelly: Thanks Rasmis, I hope to be back soon
Dell-John Kelly: Rasmus even
Rasmus Haslund: : )
Lance Boley: Thanks John for being our guest today, it was a great topic.