Anyone that has seen me present in the past two years has listened to me rant about the high cost of desktop virtualization and I did not hesitate to put the blame directly where it belonged: storage vendors. We’ve seen some crazy numbers at Gartner, upwards of $1000 per desktop for storage alone! Every time I hear that number I think of Seth and Amy on SNL doing the #Really skit.
$1000 for just the storage on 1 desktop!!!
Well the good news is that so long as you don’t blindly buy the storage some sales guy tells you to buy, chances are good you can solve VDI storage woes. In fact Citrix recently created a Citrix Ready VDI Capacity Validation Program for Storage Partners that I think has valuable information if your considering a VDI deployment.
Now I know this is a Citrix thing but here’s the deal; Desktops are desktops, storage is storage, and I don’t care if you use Citrix/VMware/Microsoft/Dell/Virtual Bridges/etc, If you are deploying thousands of virtual desktops, largely the overhead is due to Windows not due to the agent these VDI vendors run inside the VM. So with that in mind, take a look at the whitepapers they share. They provide detail for each storage vendor and they show LoginVSI tests (this is pretty much the de-facto standard in scale testing VDI environments) and they give details on the IOPS used and how much it cost.
For the lazy reader I created a comparison chart below, this is all public info that you can read yourself in the link above, I just consolidated the details I care about most, mainly price and performance.
*Please note: I did my best to make this accurate but I may have made a mistake (I’m human and this is just a blog not official Gartner research). If you see one let me know and I’ll update the graph. You can check all of these facts by reading the whitepapers yourself, if the mistake is in the whitepaper, tell the vendor. I noticed a few errors in the paper myself.
*The graph is sorted from least expensive to most expensive
Source: EMC XtremIO, FusionIO, GreenBytes, Netapp, Nexenta, Nimble Storage, PureStorage, Sanbolic, Solidfire, Tegile, Volin Memory, XIO
The Gartner Blog Network provides an opportunity for Gartner analysts to test ideas and move research forward. Because the content posted by Gartner analysts on this site does not undergo our standard editorial review, all comments or opinions expressed hereunder are those of the individual contributors and do not represent the views of Gartner, Inc. or its management.
No Atlantis Computing?
Would be nice to get Nutanix in there and see that cost as well. Good work otherwise (as usual)!
Yeah would be really interesting to see ilio atlantis there.
But if thing the license cost per desktop is already much higher than the nexena price.
It’s also kind of sad to see the greenbytes prices… i mean it’s the samezfs stack than nexenta, the just improve the dedup part…
It depends also on the Desktop Architecture what you deploy. If you have a XenDesktop Random Pooled Desktop and you have a Storage System with Inline Reduplication then you have a Reduplication of 99%. Storages allfsash storages how 3par, xtremio, Atlantis USX ….and many other what have inline dedupe can be much cheaper and also more powerful. You don’t need much capacity because all is Deduplicated
Reader beware: This is not an apples to apples benchmark – Citrix never intended it to be. The intent of these reference architectures to to provide a technical baseline that can be used to configure from. You have to read each of the reference architectures to understand what is really going on here.
Take for example a discrepancy I have been chatting with Gunnar about: not all vendors’ block sizes are the same. This has a massive impact on the IOPS listed in the table. If you casually scanned the comparison table, you wouldn’t know that. You’d likely include some arrays that shouldn’t be considered and chuck others from your short list that should be considered.
I am with Tegile, and we provided a relatively well balanced solution at 750 desktops. We also have an all-flash array, and would in no way hit $41/desktop with our all-flash solution. It is capable of far more desktops, so when normalized to 750 desktops, any all-flash array is going to look very expensive, but at large scale, an all flash array will keep amortizing $/desktop to a more attractive number while the Nexenta configuration and other lower-end arrays would have imploded by then.
So, if you are an IT manager tasked with deploying VDI on Citrix, do your homework and read each reference architecture before you pass judgement on any of these vendors listed in the table. I appreciate the summary Gunnar is shooting for here, but there aren’t just apples and oranges in here, it is a full Thanksgiving cornucopia of fruits and vegetables.
Thank you Gunnar for a great blog and putting together this summary table.
Configuring the array for specific workload is the most important part of providing the right solution. Nexenta doesn’t sell hardware, which gives their customer higher flexibility adjusting hardware resources based on actual solution requirements.
This is the biggest difference between Software Defined Storage and hardware array vendors. Software defines your storage, not the vendor, resulting in better ROI and a solution that delivers just the right mix of performance, availability and functionality.
Again, knowing the workload means knowing your customer’s needs. PVS is definitely lighter on storage than MCS and other available VDI provisioning options. Apples-to-apples comparison don’t always make sense, since not all customers want apples. If customer needs carrot, you can’t feed them an apple 🙂
Of course, do not base your decision on one paper, read each reference architecture, call your preferred vendor and make sure you get software to define your storage. Nexenta has many other test results/whitepapers recently posted on their website demonstrating increased density, performance and improved user-experience, as well as industry leading TCO.
In all transparency, I work at Nexenta, the above comments are my personal opinions and do not reflect those of Nexenta.
Contextual items that may help the readers of this blog and to digest the white papers:
– The Citrix prescription was to bring what you’d recommend for customers doing a 750 user deployment. There were no requirements other than can it pass the LoginVSI test.
– You couldn’t touch the servers, so no additional hardware or software could be changed, and that included any changes to the architecture used (i.e., PVS).
– The storage tested in this case had fundamentally one load to actually run for the test and that was to handle the write cache events. So the storage did a very basic job and didn’t hold the VDI images.
With that as the setting, we (GreenBytes) chose to bring what we would recommend for customers. This included running non-required tests to show what “could be” if we were allowed to create a best practice deployment where every user gets a persistent image.
– Doing the right thing for a 750 user deployment meant there’s no such thing as No High Availability. While this wasn’t a requirement to pass the test, it wouldn’t be production quality to leave it off. If you find it was, then take care with having a single point of failure.
– The storage consumption and performance for a “write cache” test was… well, simple. We only used 0.2% of the 2.4TB of storage presented because of the inline dedupe. The storage provided normally supports 1150 user – ALL with full persistent images and no PVS. So that’s why you’ll note we recommended an MCS deployment as a supplement.
– Why no PVS and MCS instead? It took over 3 days to provision the VMs from PVS to the 750 VDI virtual machines running on the server. An MCS delivery system would more closely mirror a thick clone deployment that would enable a PC user experience.
I welcome Version 2 of this test and what bar it creates for performance validation. Perhaps readers of this blog will send recommendations to Citrix for next gen test requirements. I made some already… like having users install applications from MS Software Center.
I just read the whitepaper from Nexenta on Citrix’s site. I find it rather disturbing the price fudging in the system configuration from Nexenta.
They list $105 per 2TB Seagate SAS 7.2k disk, that is about $80 under the current street price. They also list 8 24GB ECC DIMMs. DIMMs come in 16GB or 32GB not 24. Every item on their configuration is listed as costing less than it can be purchase for.
The server listed does not exist. Likely the model number is 6027R-E1R12L.
Citrix claims to verify the configurations, obviously not too closely.
What was the point of all the fudging? Even at true street prices or even suggest retail, the price per desktop would have still been only slightly higher.
If Nexenta justified fudging the parts list and got a way with it for this test, who’s to say these are the actually parts used?
I haven’t read the other white papers, but after the first one I’m not sure I trust any of them.
I participated as an implementation architect on one of the above efforts and I can you that there are major disparities. Some vendors provided what would be an actual system deployment (i.e. redundant, high performance, room for growth, etc.) while others provided a minimal system just enough to pass the VSI login test parameters. The take away is that these systems have been validated running Citrix products. Unfortunately, there really there is no fair comparison behind the cost numbers provided and they should be ignored for vendor to vendor comparisons. They probably are useful to generate a range of costs within the industry, ie it costs somewhere between $15 and $542 per user 🙂
To take these results and simply put them in a table as “comparable” is borderline irresponsible, particularly now that this is getting picked up by others as an authoritative “source: Gartner”.
As Rob stated, this wasn’t apples-apples, or even apples-pears. In SolidFire’s case, the capacity used for this workload was 0.2% of the storage system, and even PEAK performance was only 10% of guaranteed IOPS.
You could easily put 7500 users on the same configuration for the same price, or, more likely, put dozens of other workloads on the system system with Guaranteed QoS, rather than add “yet another island” of storage for VDI.
The test situation may not have been ideal, but I don’t fault Citrix. They aren’t the one putting all the results in a table that implies they are comparable – because they aren’t. They simply post the whitepapers so that you get the full context of each test.
Also, fwiw, in addition to the questionable pricing on the “lowest priced” result, it also isn’t an HA configuration. Anyone recommending a non-HA storage system for a 750 user VDI workload needs to get their head examined.
As a technical representative of X-IO that participated in the testing, this was an excellent opportunity to directly interface with some of the best XenDesktop performance people in Citrix. . We are grateful that Citrix invited us to be a part of this initial round of testing. It’s clear however that some key points need to be emphasized here:
• The costs provided are simply misleading as warranty information is not included in most of the submissions. Warranty costs can add 70% (and more) to the overall solution cost used to compute $/Desktop. Without this information, there is not an accurate view of the total solution cost.
• The capacities reported by the vendors are estimates only. There was no actual data that was stored for 87% of the capacity (so these numbers are completely based on vendors “best case” estimates and we all know the difference between “marketing” and “real-world” statistics).
• Their “Estimates” for max numbers of VDI users are wrong. They’ve based the IOPS on synthetic load generators (so not real data). They are quite simply trying to justify not being able to size this “small”.
These, and more, require that each submission be carefully evaluated. Remember, if it looks too good to be true, it probably isn’t….
Warranty costs can comprise 10%/yr (and higher) of the total solution costs in the first 3 years (30%), and 20%/yr in 4 and 5 (40%). So these can make up a significant cost of the solution, and only 4 out of 12 vendors reported any warranty information included in their reported “cost”. Here’s the breakdown:
• Nimble: 1yr, Next Business Day
• Nexenta: 1yr “Gold” support (you have to add the support of the physical server to this)
• NetApp: 3yrs, Next Business Day
• X-IO, 5-yrs, 4-Hr Parts On-Site
The other 75% of the submissions don’t talk about the warranty at all – I wonder why that is?
As a baseline for design, not everything can be tested that will have an impact to the end user. While allocated space for user data was required, no actual data was put onto the storage system in this test. So most of the capacity (user data, 87% of total) is based on the “estimated” compression/deduplication/etc” ratio (11/12 submissions) of each vendor. If the user data doesn’t compress to this ratio, you are looking at more hardware required (there are pretty aggressive ratios here). Some architectures have performance problems when they have to do this over a lot of data (as would be the case here), and would likely require more equipment as well. (Full Disclosure: X-IO does not do this by design).
Then (to make matters worse), some vendors report benchmark numbers in their papers for “estimates” (the really big IOPS numbers in the article’s table). What did those have to do with the tests run here? Absolutely nothing, and representing this as something that customers can reasonably expect in a VDI environment falls somewhere between specious and delusional.
While some sized for much larger numbers of desktops (overprovisioned for performance), we chose an ISE design that satisfied both the performance and capacity requirements of the test (LoginVSI Score 1,211) for the least TCO over 5-years. Could we have used our Hybrid array? Sure, but it had more performance than this test required, and would have increased the cost unnecessarily.
So was this test perfect? No, and not any test ever will be as long as there are vendors (we all have different viewpoints and goals). I feel they more prove that vendors are willing to commit money/resources/knowledge/people to further the understanding of the different parts to the storage “value” discussion. We are looking forward to the next round of testing and for there to be much more truth and transparency in the storage industry.
Chiming along with others here, the Citrix capacity program hosted just 750 desktops and Pure Storage FA-420 consumed 1% of the storage and had hardly consumed 1% of the IOPS our AFA can offer. A properly configured Pure FA-420 can host upto 5000 desktops and allows you to run other workloads if you wish. So if we do the math you can get cost of desktop way lower.
Customer running Pure already realize this and getting best user response time and saving ton of money on storage costs.
There are real world scenarios like what if 500 users logged out at 5:00 PM and logged in at 8:00 AM, does the storage withstand the log-out and login storms?
How fast can I create a pool of desktops? How efficiently I can patch the desktops?
How can I withstand the virus scans and can I have virus scanning as part of the image?
These are real life problems, that can only be addressed by real storage solution and Flash by far has answers to all the problems.
Pure customers have deployed 100s of thousands of desktops till date and realizing a fresh thinking is needed to address all the above mentioned problems.
I don’t want to poke holes (not our style) in the lowest priced storage but I would encourage readers to read carefully and see for themselves if they really want to implement their desktop solutions on such a storage solution.
Data Center Architect @PureStorage