Atlantis adds persistent support to ILIO Diskess VDI, creates the ultimate VDI storage solution

Last September Atlantis Computing released a product called "ILIO Diskless VDI." I wrote about it on BrianMadden.

Last September Atlantis Computing released a product called "ILIO Diskless VDI." I wrote about it on then, and Gabe and I gave it the "Best of VMworld 2012" award in the desktop virtualization category. If you're not familiar with ILIO Diskless VDI, the basic idea is that you run an Atlantis storage appliance VM on each of your VDI hosts and it uses RAM to store all the deduped, compressed disk blocks of all the desktop VMs on that host. So essentially you're using a RAM disk for the C: drive of each VDI instance, meaning they run fast. Really fast. Screaming crazy fast. And since Atlantis is only storing each disk block once (regardless of how many VMs on that host use that block), they can run dozens or hundreds of VMs off of a very small amount of memory.

The only problem with the Diskless VDI solution was that it only worked with non-persistent "shared" VDI (where all users share a golden master disk image). As you know from my article yesterday (and various other rants over the years), I believe that VDI really shines with persistent "one-to-one"desktops since (1) that's the way your current pre-VDI physical desktops work, and (2) if your desktop requirements were simple enough that gold master image sharing worked, why wouldn't you just use RDSH? (Though true there are reasons that you'd want to use a shared master, especially with something like Unidesk or Mirage, but my "default" is to do VDI just like you do physical, and that's 1-to-1.)

So while Atlantis's ILIO Diskless VDI of 2012 was cool, it wasn't useful enough to break out since it only worked with shared images.

But that all changes today, as Atlantis has released a new version of their ILIO Diskless VDI for persistent desktops. I can safely say that this product seems like the #1 ultimate best VDI storage solution on the market today. Period.

How it works

From an architecture standpoint, Atlantis's ILIO Persistent VDI product has a VM on each VDI server that they call a "Session Host." That's really no different than their previous non-persistent product. The new part is that they now also have a "Replication Host" which is a central VM that maintains the master copy of each user's blocks. (And of course this can be replicated, synced, backed up, etc.) So when a user wants to connect to their desktop, the VDI software (View, XenDesktop, etc.) routes them to some random VDI server, and that server's Atlantis ILIO Session Host VM gets all that user's blocks it doesn't already have (from other VMs) copied down to its memory from the central Replication Host.

The beauty of this is that the long-term permanent storage of each user's unique blocks is outside of the dekstop VM's IO data path. The desktop VM on the VDI host reads and writes its blocks from the RAM of the ILIO Session Host VM at full RAM speed, and the Session Host just replicates any unique compressed blocks over to the Replication Host out-of-band, taking its own sweet time. (Which, in disk speed land, is still within a few seconds.) This dual-role architecture is how Atlantis solves the CAP theorem. (Do you know of the CAP theorem? It says that when it comes to storage, you have Consistency, Availability, and Partition Tolerance—often manifested as lots of local data partitions which means the "P" ends up being "Performance"—and you can only get two out of the three out of a single system.) A lot of vendors opt for Consistency and Availability and then throw a bunch of SSDs or DRAM at the thing to solve the "Performance" part, but Altantis get the Consistency and Performance from the local Session Host and the Consistency and Availability of the centralized Replication Host. The Availability portion of the Replication Host is pretty obvious, but when it comes to consistency, Atlantis is able to keep the central host up-to-date pretty easily since they only have to write blocks that change or track small metadata, and because they're deduping everything they only have to send blocks that are new that the central Replication Host hasn't ever seen. And of course those blocks are all compressed, they serialize the writes, and they combine small changes into single blocks, so really you can imagine that hundreds of desktops can use a low-end NAS as their primary storage.

Check out this crazy amazing performance

Do you remember Chetan Venkatesh's (Atlantis CTO and Founder) session from BriForum 2010? He gave a talk where he basically said that in the future, desktops running on VDI will actually be faster than desktops running on local laptops. He based his argument on Moore's Law and shared resources and stuff like that, and his point was that some day soon, everything that a VDI desktop runs will be more-or-less instant. That presentation was over two years ago and I don't know if Chetan knew about what he wanted to do with this Diskless VDI stuff, but let me tell you, Atlantis Computing has built that. They gave me a persistent VM—complete with admin rights and everything—in their in-house production XenDesktop environment which is powered by ILIO Persistent Diskless VDI. In other words, I was running a dekstop alongside of Chetan and Bernard and Jim—just like an Atlantis employee. Since this was a persistent desktop I was able to install Chrome, Office 2013, Dropbox, 1Password, and all my other applications. And let me tell you, everything in this dekstop was instant... just so fast! I opened my huge 200-slide 100MB PowerPoint presentation and it happens instantly. Seriously, I couldn't even time it with a stopwatch. I double-click on the .pptx file and the PowerPoint splash screen is up for about 1/2 second and then my full presentation is loaded. The same goes for Word, Excel, Chrome.. everything just feels instant. This is absolutely the fastest desktop I've ever used in my life.

To prove my point, I installed the PassMark PerformanceTest 8.0 benchmarking suite into my VM and ran the DiskMark disk benchmarking utility. Check out Perfmon during that time:

ILIO 4 IOPS during disk test

Notice that for this graph of the IOPS I set the scale to 0.1, and this thing was able to deliver a maximum of over 30,000 IOPS! How crazy is that? This VM is pretty standard otherwise: Windows 7 x64, 2GB RAM (1.25GB physical with 750MB overcommit), 40GB disk size. If the peaks of 30k IOPS aren't enough to blow you away, check out the overall benchmark scores that the PassMark app showed when it was done. (The green line is me, and the red lines are other common disks that were tested):

ILIO 4 disk performance


This is crazy. My overall Disk Mark score was 4355, and if you look at the PassMark results of the current fastest drives (which is updated from thousands of tests daily), this puts my Atlantis ILIO drive in the #7 position. Overall. Of all the drives in the world. Faster than most SSDs.

This thing is crazy fast (which makes sense again since all the inline IO is in RAM), and it's a full real persistent desktop.

And remember, this Atlantis ILIO Persistent VDI is pure software

In addition to the crazy performance, remember this Atlantis ILIO solution is pure software. That's awesome for a bunch of reasons. First, it means you can scale as needed. It's not like you have to buy a $90k appliance that you then have to use for at least 2,000 users to get your money's worth. If you only have 20 users then you can buy 20 licenses (estimated around $150 each, depending on your reseller).

Being all software also means that you don't have to change architectures part way through your project. The ILIO stuff is easy enough to buy for a tiny environment and scalable enough for thousands of users. Also since it's all software, you don't need to use a PCI slot for Fusion-IO or anything, so you can pop an Nvidia GPU or Teradici card in there to get some great graphical performance.

And finally, being all software means that you don't need to buy any fancy high-speed storage to make it work. Even your old server with dual hard drives from 2005 can handle it. No SSD. No Violin. No Fusion-IO. No storage appliance. Etc.

Details of the test environment

Here's a look at the details of the VDI environment I used for the past month. Notice that there's nothing really impressive here at all. The individual machines and primary storage are not that special.

Storage Specs

  • NetApp Aggregate in RAID-DP with 10x 15k RPM SAS drive. (180GB each / 1.8TB RAW space). They tested this to handle 800 users in the LoginVSI Heavy test. This disk is only being used for replication, backup and recovery in case of a host failure.
  • 1 NetApp Volume exported as NFS to the ILIO Replication Host over 2x 1Gbe (for traffic path redundancy)

Server Specs

  • 12 cores (Dual hex-core) 2.93Ghz CPUs
  • 144GB RRD3 RAM
  • No local storage for VMs. (Only the hypervisor is installed on local 15K SAS disk storage)
  • 3x 1Gbe bonded for storage traffic
  • 1x 1Gbe for management


  • VMware vSphere 5.0 and Virtual Center 5.0
  • Citrix XenDesktop 5.6
  • ILIO persistent VDI 4.0

Desktop &  Configuration

  • 50 Windows 7 VDI VMs
  • RAM per desktop is 1.25GB with an overcommit of 750MB. (So each desktop VM thinks it has 2GB RAM.)
  • The ILIO Session Host VM has 75GB of RAM to use for internal OS and RAM storage

RAM and Storage Consumption Stats

  • On my XenDesktop server, the Atlantis ILIO Persistent VDI is turning its 75GB of physical RAM into 1.36TB of virtual NFS storage.
  • On a per desktop basis, this means that each VM is using 28GB on average (40GB allocated) of virtual storage, but that is consuming only 1.5GB of physical RAM per desktop. My desktop, after a month of use, is right about at average using 27.7GB of virtual storage.
  • On the NetApp backend storage, the whole VDI environment is using 98GB (~1.9GB of storage per VM).
  • With 1.25GB of RAM used per desktop VM and 1.5GB of RAM used for the ILIO Session Host, I'm only using a total of 2.75GB of RAM for my desktop VM memory and storage combined. They can overcommit more agressively when using RAM as storage because the paging penalty is negligle (since they're paging onto ILIO RAM).


Like I said, Atlantis ILIO Persistent diskless VDI is the current storage solution to beat. You have a crazy amount of IOPS with full persistence of disk images in a software-only solution that doesn't require any special hardware for about $150 per user. It's amazing, and I'm hooked. What do you think?

Join the conversation


Send me notifications when other members comment.

Please create a username to comment.

Now they have conquered the persistent desktop, is there anything stopping Atlantis moving into the server market? I would love to implement this in both my VDI and server clusters!


The replication host is definitely the missing piece in the Atlantis ILIO puzzle we all been waiting for (maybe not all, but at least I did).

I wonder how scalable the replication host is and what the implications are on HA and DR scenarios. My first imperssion is that this piece makes implementing a scalable and high available environment incuding Atlants ILIO more realistic (compared to ILIO 3).

Do you know it there are plans to use the same architecture for the XenApp product?


BTW, do you know if you can now mix persistent and non-persistent VDI's using the same ILIO appliance?

That would make it the best next thing since sliced bread!


Thanks for your encouraging responses. Here are some answers to your questions:

• Replication Host Scalability – The replication host can handle 500-1,000 persistent virtual desktops depending on the workload. We have a LoginVSI Heavy test of 800 users on a single replication host with a VSI Max Not Reached and a baseline of 992ms response time.  

• XenApp Architecture – You can use our ILIO for XenApp Product with the same In-Memory technology for Citrix XenApp deployments with the PVS write caches on the Atlantis ILIO datastore.  

• HA and Disaster Recovery – Both Atlantis ILIO 3.2 and Atlantis ILIO 4.0 fully integrate with HA and DRS – we have many customers using this where availability of the desktops is critical.  The difference in Atlantis ILIO 4.0 is that you can use local RAM as primary storage, while still providing the ability to use HA and DRS. And for DR protection of persistent deployments, ILIO reduces the data that needs to be replicated by the storage systems very significantly and has deployment, configuration and provisioning automation, giving you much lower RPO and RTO! We have more in the works on DR. So, stay tuned.

• Stateless – You can use the persistent architecture for non-persistent as well. This might make sense for a mixed environment. However, in a non-persistent VDI or XenApp deployment, you don’t really need the replication host since HA is provided at the broker level.  So, most customers will separate their persistent and non-persistent desktops onto separate servers to minimize hardware but you can support a mixed environment.  

• Server Market – Yes, yes, yes! Atlantis ILIO for Servers - stay tuned!


Got to say that minus the fact that I haven't had my hands on it yet personally, this could be the biggest thing I have seen in a while for the VDI space. IOPS or the lack there of in most environments is the biggest hurdle to overcome. Can't wait to get this in my hands.


Hi Mark,

PVS can use RAM for write cache as is, what are the benefits of using ILIO?

I suppose that ILIO would need less RAM as it consolidates and de-duplicates writes from all of the VMs but will this benefit alone will cover the license cost?

Staying tuned for ILIO for servers (FlexCloud is it?)!!


Hi Ron,

While PVS can use server RAM as the location for the PVS Write Cache, it does not optimize the use of server memory at all.  For a host with 100 virtual desktops, each with a 6GB write Cache, you would need ~800GB of server memory.  (Virtual desktop memory = 2GB, Hypervisor memory 6GB)

Atlantis ILIO consolidates, deduplicates and compresses the Write Cache data, typically reducing the capacity by 90%.  For the same 100 desktops, you would only need 272GB of RAM with Atlantis ILIO.  Or in other words, you could fit ~93VMs on a blade with 256 GB of memory.

Calculating with list prices for UCS memory (scaled down density to 93 VMs to fit in 768GB as I don't think any of the UCS blades can scale beyond 768GB) and Atlantis ILIO, you would save ~70% of the memory increase purchase if you went with Atlantis ILIO

Also, I would like to refer you to the Reference Architecture we did together with Citrix and GlassHouse technologies last year.  It's done with MCS rather than PVS, but the ILIO per desktop memory overhead was only 139MB!


-- Mark


Thanks for the information, it's really rocking!

Though I was referring to the XenApp use case.

For VDI I think this is great, especially now  with persistent VDI


Hi Mark,

Does the ILIO 4 work successfully with Citrix Personal vDisk where you have a shared Citrix PVS based vdisk shared between all users, a personal vdisk for each user and a write cache for eachl users?

Can ILIO 4 consolidate, dedup and compress on the users personal vdisk as well as the PVS vdisk and the users write cache?

Does it actually remove the need for Citrix Personal vDisk because of the storage saved by using ILIO 4?




Hi Mark,

  Another question for you:-

 Are XenServer and Hyper-V also supported as hypervisors?




Enterprises will definitely go for this.

Would it be better for SMB's go for this approach along with XenDesktop/View and reduce storage costs or go for solutions like VDI-in-a-box and reduce overall cost per desktop?


Atlantis always rocked. They just continue to rock more..  I simply can't imagine why anyone would do VDI any other way (unless their storage vendor is taking them on a lot of golf outings....


Ron, the same holds true for XenApp and PVS – a typical XenApp hosted desktop generates about 8 IOPS per user.  While this is much lower than a typical VDI workload, at scale it still is very expensive to deliver this with traditional storage.   We have a customer that is using 50 GB Write Caches for his XenApp virtual machines.  With ILIO that reduces down to about 5 GB per Virtual Machine, so using memory is actually feasible.   Additional benefit is that you can oversize your Write Caches quite significantly, so you never run the risk of blue screening when you run out of Write Cache capacity, and you eliminate any dependency on shared storage infrastructure for your XenApp machines.

Shaun – Atlantis ILIO fully supports Personal vDisks along with all flavors of virtual desktops, whether it is with PVS or MCS. I don’t have concrete test data at hand about optimization rates for Personal vDisks, but I suspect the optimization rates are high.  

I would like to point out one benefit of using ILIO in combination with Personal vDisk: since the PVD’s get attached at boot rather than at logon, the session login times can be high if the base image is on slow storage. Since virtual desktops on Atlantis boot in as little as 12 seconds, you can speed up your login times significantly.  

And yes, XenServer and Hyper-V are supported hypervisor platforms.

Prudhvi – we have customers with as little as 25 desktops.  Our license model is per-user, so it scales with the size of your deployment, and installation and configuration is automated for a very userfriendly management experience.

Thanks for the shout-out, Jan!  Keep the questions coming!


Hello maybe i don't understand something here, but if you are doing persistent desktop you can't only use ram. At some point you must have a disk somewhere (if don't you will always have a risk to loose data).

I mean in persistent mode your vscaler just take the IO and will make deduplication on ram (agree with that), it will change the IO pattern because it's a transactional filesystem (so it kept something like 10 s in ram, and try to transform the random pattern into sequential pattern).

The thing here is if you have a power loss and don't copy this 10 s on a disk, you loose those 10 s (that's why other transactional file system use nvram or ssd to store the journal of this 10 s)

And finally the compression is not make by the filesystem itself but by the linux kernel module zram (lzo compression),

If i'm wrong in my statement please explain me.



@Aglidic, correct, you can't just use RAM. That's where the replication host comes in. (Re-read the section on "how it works" for the complete description.) True, there might be some delay in writing those blocks to disk, but if you have a hard power fail of a persistent VDI environment (which is unlikely in a datacenter), then you're screwed regardless of whether or not you're writing directly to disk or on a 10s delay.


Ok thanks for the anser it makes more sens now.

So in that case the Ilio controller are juste big caching machine with dedulication.

And by the way if you have the information that would be great, what hashing algorithm did they use for deduplication?

I mean i create a zfs box to make a little comparison, i know than zfs use sha-256 for dedup and in my zfs box in need 4x cpu power to keep the same performance than the ilio. (so ok i didn't made optimization or stuff like that but still)