Sizing your Cisco HyperFlex Cluster


Update 31st July 2018

Since my post below, Cisco have released a Hyperflex profiler and sizer tool, for more info of where to get it and how to install and use it, refer to the below video by Joost van der Made.


But still feel free to read my original post below, as it will teach you the under the covers Math.

/End of update


Original Post:

Calculating the usable capacity of your HyperFlex cluster under all scenarios is worth taking some time to fully understand

The most significant design decision to take, is whether to optimise your HyperFlex cluster for Capacity or Availability, as with most things in life there is a trade-off to be made.

If you size your cluster for maximum capacity obviously you will lose some availability, likewise if you optimise your cluster for availability you will not have as much usable capacity available to you.

The setting that determines this is the Replication Factor

As the replication factor is set at cluster creation time, and cannot be changed once set , it is worth thinking about. In reality you may decide to choose a different replication factor based on the cluster use case, i.e. Availability optimised for production and capacity optimised for Test/Dev for example.

It is absolutely fine to have multiple clusters with different RF settings within the same UCS Domain.

So let’s look at the two replication factors available:

Replication Factors

As you can no doubt determine from the above, the more nodes you have in your cluster the more available it becomes, and able to withstand multiple failure points. In short the more hosts and disks there are on which to stripe the data across the less likely it is that multiple failures will effect replicas of the same data.



There is 8% capacity that needs to be factored in for the Meta data ENOSPC* buffer

*No space available for writing onto the storage device


Now for some good news, what the Availability Gods taketh away, HyperFlex can giveth back! remember that “always on, nothing to configure, minimal performance impacting” Deduplication and Compression I mentioned in my last post? Well it really makes great use of that remaining available capacity.

VDI VM’s will practically take up no space at all, and will give circa 95% capacity savings. For persistent user data and general Virtual Server Infrastructure (VSI) VM’s the capacity savings are still a not to be sniffed at 20-50%.


The last figure to bear in mind is our old friend N+1, so factor in enough capacity to cope with a single node failure, planned or otherwise, while still maintaining your minimum capacity requirement.


As in most cases, it may be clearer if I give a quick manual example.

I’m only going to base this example on the storage requirement,  as the main point of this post is to show how the Replication Factor (RF) affects capacity and availability. Obviously in the real world vCPU to Physical Core ratios and vRAM requirements would also be worked out and factored in.

But let’s keep the numbers nice and simple and again not necessarily “real world” and plan for an infrastructure to support 500 VM’s each requiring 50GB of storage.

So 500 x 50GB gives us a requirement of 25TB capacity.

So let’s say we spec up 5 x HyperFlex Nodes each contributing 10TB RAW capacity to the cluster which gives us 50 TB RAW, but let’s allow for that 8% of Meta Data overhead (4TB)

50TB – 4TB = 46TB

So we are down to 46TB we now need to consider our Replication Factor (RF), so let’s say we leave it at the default of RF 3, so we need to divide our remaining capacity by our RF, in this case 3, to allow for the 2 additional copies of all data blocks.

46TB / 3 = 15.33TB

So that puts us down to 15.33TB actual usable capacity in the cluster.

So remember that 25TB of capacity we need for our 500 Virtual Servers, well lets be optimistic and assume we will get the upper end of our Dedupe and Compression savings to lets reduce that by 50%

25TB / 2 = 12.5TB

And that’s not even taking into account thin provisioning which although variable and dependant on a particular clients comfort factor would realistically reduce this capacity requirement by a further 30-50%

So let’s now say our 12.5TB on disk capacity is realistically likely to be 6.25 -8.75 TB

meaning we are OK with our available 15.33TB on our 5 node cluster.

So the last consideration is could we withstand a single node failure or planned host upgrade and still meet our minimum storage requirement?

So either we could work out all the above again but just using 4 nodes (contributing a total of 40TB RAW) to the cluster. Or, since we know that 5 nodes contribute a total of 15.33TB usable, then we know that each node is contributing approx. 3.06TB of usable capacity, so if we take a node out of the cluster.

15.33TB – 3.06TB = 12.27TB usable remaining, which is still above our 8.75TB realistic requirement.

Luckily to make all this sizing a walk in the park, a HyperFlex sizing calculator will soon be available on CCO.



About ucsguru

Principal Consultant and Data Center Subject Matter Expert. I do not work or speak for Cisco or any other vendor.
This entry was posted in HyperFlex and tagged , , , , , , , , . Bookmark the permalink.

5 Responses to Sizing your Cisco HyperFlex Cluster

  1. Daniel Harf says:

    Thanks mate!

  2. Drunken_Commie says:

    You probably had it right the first time 😉 You need 8% for ENOSPCE (‘no space left on device’ – important due to ‘constantly keep adding’ feature of log structured file system) and you need _extra_ space for metadata (‘where’s my data, what it is, what references it’ etc). Metadata starts small but grows with time (when you have more data, more clones, snapshots, and so on) and can reach even 10%.
    Hence the “leave 20% space empty”.

  3. Pingback: Sizing your Cisco HyperFlex Cluster | | IT-SERVICES

  4. Romel says:

    can VM in hyperflex cluster connect to external storage resource, either via nfs or iscsi?

    • ucsguru says:

      Absolutely! But obviously VM’s outside of the HyperFlex environment can’t access the CiscoHX storage. However Compute only blades can. (See Hybrid Cluster)

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.