«

»

Nov 30 2011

Use Flash Drives (SSD) for Linked Clones, not Replicas

Advertisement

In my previous article EMC FAST Cache effectiveness with VDI I focused on FAST Cache benefits for VDI deployments. The tests I completed also generated a huge quantity of data that is tremendously valuable for better sizing of VDI environments.

I previously have discussed (here) how IOs suffer severe split between replica, linked clone and persistent disks. Depending on the storage tier serving the data it is possible to have different latencies for different parts of the Guest OS (Windows XP or Windows 7). However, in a linked clones environment the disks are interdependent at GuestOS layer, and a VMDK suffering from latency will also affect a different VMDK that is not suffering. Example: Replica disks that are backed by Flash Drive drives with Linked Clones that are backed by SATA drives.

In my validation with 200 VMs during Steady State each VM has an average of 8 IOPs. The 8 IOPS per virtual desktop is split into 0.29 IOPs for the Replica and 7 IOPs for the Linked Clone.

This represents 4% read-only operations for the replica and 96% read/write operations for the Linked Clone.

These numbers clearly identity that during Steady State the Replica disks are barely used and most of the IOs are absorbed by the linked clones. The two graphs below demonstrate the IO split between replicas and linked clones.

 

clip_image002

 

With knowledge of the disk behavior it is possible to size storage for the exact performance required in each tier.

VMware recommend placing Replicas disks in a Flash Drive (SSD) tier. However, you know now that Flash drives would only really be effectively used out of the Steady State period.

On the right picture (above) two spikes are clearly observed. The first is the logon storm; the second is the VMware View Refresh operation storm. Those are the periods when the Replica disks are utilized more aggressively. However, Linked Clones are equally utilized.

In fact, Linked Clones disks are more comprehensively utilized than the Replicas. This actually makes me evidently think that if there is a limited number of Flash Drives to be utilized they should be supporting Linked Clones, not Replicas. Most recent intelligent storage arrays will cache most utilized blocks in L1 DRAM cache, and if the Replica disk is being heavily utilized those blocks will end up in cache anyway.

It all comes down to how the VDI environment is utilized by the users. If boot storms, logon storms and VM refreshes are often requested, such in a Non-Persistent Pool serving school classes, then it is important to support Replica disks as much as possible. That means Replica disks should be on the fastest storage tier.

If there are no constant VM refresh or extreme login storms, then it is probably it is better to support Linked Clones as much as possible. That means Linked Clones should be on the fastest storage tier.

 

I know, I know… that goes against industry recommendation. Well, I guess there is always someone to go 1st against the flow. I have done my tests and math and I recommend you to do yours.

Of course, if your organization have budget to employ Flash Drives for everything, go for it! There are technologies on the market that help in reducing the number of disks required to serve a larger number of IOPs without having to utilize an all SSD based array. Find out more here.

 

Similar Posts:

Permanent link to this article: http://myvirtualcloud.net/?p=2513

5 comments

4 pings

Skip to comment form

  1. Barrie

    Good article Andre – this is mandatory knowledge for anyone implementing a linked clone View environment.

    For quite some time now I have used local SSD in the ESXi hosts for replicas and linked clones – VDI is not cost effective if trying to do all this with uber $$ tiered shared storage.

    Atlantis et al don’t appear to alleviate the cost at $150 US per user – again, more expensive than good Intel MLC SSD locally which is very cheap now and can handle around 5-6 years of continuous VDI write use according to some VMworld preso’s

    Storage is the most expensive part of VDI if not done correctly and if so, XenApp and Terminal servers might be cheaper and more effective.

    This is my 2c – people may have other opinions :)

  2. Andre Leibovici

    @Barrie
    Thank for your comments.

    I look at VDI as a multidisciplinary solution that may require different techniques and optimizations according to the end-goal. There are scenarios where only SSD make sense, however there are other scenarios where auto-tiering solutions will make more sense.

    Andre

  3. Mac Binesh

    Hi Andre,

    Interesting article. I do agree with your conclusions. I did published View 4.5 stateless desktops with local SSD with both replica and linked clonses on Intel MLC 160 GB SSD drives. We are updating that with Intel 300 GB drives to be published by PEX 2012 – Stay tuned.

    Here is the link to the View 4.5 Reference Architecture stateless desktop on local SSD:
    http://www.vmware.com/files/pdf/VMware-View-45-Stateless-RA.pdf

    Mac Binesh | Group Product Manager | End User Computing | VMware Inc.
    mbinesh@vmware.com

  4. Dan Brinkmann

    I would be curious if this IO pattern is also consistent when using Citrix MCS provisioning of desktops, it’s a similar architecture and I can’t think of any reason the IO patterns would be different than what you broke out here.

  5. Andre Leibovici

    @Dan Brinkmann
    The IO pattern for Citrix MCS should be the same since this is actually driven by Windows. However, MCS does not have support for single replica datastore and each datastore will have 1 or more replica. From a sizing perspective this will completely change how you should make the calculations for a properly architected VDI environment.

    Thanks for your comment,
    Andre

  1. Technology Short Take #18 - blog.scottlowe.org - The weblog of an IT pro specializing in virtualization, storage, and servers

    [...] Leibovici bucks the “common wisdom” and recommends that EFDs should be used for the linked clones, not the base replica. I can’t really argue with his recommendation; it’s based on a pretty thorough analysis [...]

  2. NetApp FlashCache Considerations in a VMware View Environment « Wahl Network

    [...] to pause and point out an article written by a vSpecialist named Andre Leibovici entitled ”Use Flash Drives (SSD) for Linked Clones, not Replicas” that demonstrates that perhaps replicas aren’t so deserving of tier zero storage after [...]

  3. Environmentalism World » Blog Archive » NetApp FlashCache Considerations in a VMware View Environment …

    [...] here to pause and point out an article written by a vSpecialist named Andre Leibovici entitled ”Use Flash Drives (SSD) for Linked Clones, not Replicas” that demonstrates that perhaps replicas aren’t so deserving of tier zero storage after all [...]

Leave a Reply