Yellow Bricks

by Duncan Epping

Order of storage tiers… (via twitter @mike_laverick)

· ·

@Mike_Laverick asked a question on twitter today about something that is stated in the Cloud Computing with vCloud Director book. His question was, and no he is not dyslectic he only had 140 characters 🙂

pg65. Order of storage tiers. Doesn’t that infer FC/SDD+VMFS is “race horse” and NFS “donkey”…???

Mike was referring to the following section in the book:

SLA Service Cost RTO Storage RAID Applications
Tier 0 Premium $$$$$ 20 min SSD, FC 1+0 Exchange, SQL
Tier 1 Enterprise $$$$ 1 hour FC 1+0, 5 Web servers, Sharepoint
Tier 2 Professional $$$ 2 hours iSCSI, NFS 3, 5, X Custom apps, QA
Tier 3 Basic $ 2 days NFS 3, 5, X Dev/Test

This basically states, as Mike elegantly translated, that FC/SSD is top performing storage while NFS is slow or should I say “donkey”. Mike’s comment is completely fair. I don’t agree with this table and actually did recommend changing it, somehow that got lost during the editing phase. In the first place we shouldn’t have mixed protocols with disks. Even an FC array will perform crap if you have SATA spindles backing your VMFS volumes. Secondly, there is no way you could compare these really as there are so many factors to take in to account ranging from cache to raid-level to wire speed. I guess it is still an example as clearly mentioned on page 64, nevertheless it is misleading. I would personally prefer to have listed it as follows:

SLA Service Cost RTO Protocol Disk RAID BC/DR
Tier 1 Enterprise $$$ 20 min FC 8GBps SSD 10 Sync replication
Tier 2 Professional $$ 1 hour NFS 10GBps FC 15k 6 Async Replication
Tier 3 Basic $ 1 day iSCSI 1GBps SATA 7k 5 Backup

Of course with the side note that performance is not solely dictated by the transport mechanism used, there is no reason why NFS couldn’t or shouldn’t be Tier 1 to be honest. Once again this is just an example. Thanks Mike for pointing it out,

Reader Interactions

Comments

  1. Domenico Viggiani says

    Why do you say that “Even an FC array will perform crap if you have SATA spindles backing your VMFS volumes”.
    Mixing SATA/FC/SSD disks in a Storage Pool is pretty common on EMC “unified” arrays with auto-tiering active (+ FastCache).

    I agree with you that “NFS” is not strictly a tier.

    • Aaron K says

      Another comment on the first diagram, is that with the improvements to Exchange over the years (especially Exchange 2010), it can be relegated to non-replicated, low/IOPS disk (SATA even) when used with DAGS.

    • Duncan Epping says

      Because performance of an FC array backed by SATA drives is crap. I did not mention FAST or any other tiering solution out there did I? Of course I understand that with auto-tiering solutions the experience might be different. Note I used might as there is no guarantee your blocks will be on fast storage when you start pushing the limits!

  2. Prasenjit says

    Duncan,

    Can’t we consider iSCSI as Tier 1 or Tier 2 even. Since a long time iSCSI became very robust and becoming more robust with 10GB line speed, iSCSI Multipathing capabilities and high speed spindles.

    SLA Service Cost RTO Protocol Disk RAID BC/DR
    Tier 1 Enterprise $$$ 20 min FC 8GBps SSD 10 Sync replication
    Tier 2 Professional $$ 1 hour iSCSI 10GBps SAS 15k 5 Multisite Sync Replication

      • Carl Skow says

        I respectfully disagree with this statement. There is no hard evidence to support that the transport is truly relevant in the age of gigabit Ethernet and vSphere 4.1.

        You recently posted the new TR by NetApp (http://media.netapp.com/documents/tr-3916.pdf) that shows that under certain workloads there is a 4-6% variation between 8Gb FC and 1 Gb iSCSI. While academically that is important, in terms of presenting to customers/business units it is much much less important than the other relevant attributes of storage (disk type, RAID, replication, array).

        • Duncan Epping says

          I know, but once again it will depend on the exact config. When you have an array with 3 x 1GBps iSCSI or an array with 2 8GBps ports, I am sure you will reach the limit sooner or later on both and as such it will be part of the tier. maybe not from a 1:1 performance perspective when you run 10 vms, but also think about upfront costs / pricing etc.

  3. Carl Skow says

    I can’t believe the transport is being viewed as an important facet of the “tier”… It’s about the quality and speed of the spindles/SSD behind it, as well as the RAID type, as well as replication.

    • Duncan Epping says

      Well clearly there is a huge difference between 1GBps iSCSI and 8GBps FC so it should be considered. However I am just giving an example of how you could set this up. It is not a recommendation or best practice,

      • Mxx says

        or 4GBps FC vs 10GBps iSCSI 🙂

        I’m sorry for a stupid question but what exactly would you use such a table for?
        It seems like there are a lot more “it depends” caveats than ‘rules of thumb’.

        The only people I see using this kind of breakdown would be a storage vendor that tries to muddy up buyer’s understanding of different technologies…

        • Duncan Epping says

          It was just an example of how different tiers of storage could be offered in a vCloud environment. As said, nothing more than just an example but in the case of the first one just poorly chosen.

  4. Julian Wood says

    Disk speed is the primary thing you need to factor in, not protocol. If you have high IO VMs you need faster disks (if your IO is primarily reads then some storage caching may help though).
    Your disk IO normally always gets saturated way before your transport mechanism.
    Use NFS over 10GbE and simplify your life!
    No LUNs, initiator groups, masking, expensive FC switches.
    Rather have fewer, simpler, bigger, more flexible datastores.
    Move to NFS and you won’t look back other than to reminisce over how much hassle you used to have with block based storage.

    • Duncan Epping says

      It about the Service you are offering and the way it is positioned. I understand that the disks will largely drive the performance, but more importantly for a service like this is the reliability and of course the cost associated with it. The example mere shows one of the options you have. You could leave columns out if you feel that fits your situation better.

  5. Andy says

    Thanks for the clarification Duncan. I was irked when I first read that in the book as we use iSCSI for tier 1 per our design and we see no issues in meeting and exceeding our requirements.

    The transport medium is important but not as important as all of the other factors IMO.