Total Pageviews

Thursday, October 4, 2012

Workload Placement and the economics of Cloud

Had a very interesting meeting with a global outsourcer yesterday where got under the covers of their rate card for private cloud so thought I would share some observations.

My presentation on workload characteristics was entirely focused on how all workloads are different and therefore require different platforms.  Whilst presenting, this point was widely acknowledged and accepted.  Then when we get into the general discussion about the suitability of z as a cloud platform the discussion focused on how can IBM make z just like other platforms so it is easy to size. 

Now I fully understand that z is a different religion and understanding how our I/O based approach relates to cloud is not easy to articulate or fully understand from an x86 standpoint, but when you acknowledge that all workloads are different can you not acknowledge that platforms differ to?

An analogy I used after the meeting was:

Imagine you commute to work by train and a colleague commutes to work by bike, which form of transport is better?  Obviously a lot of factors come into play, some of which could be:

  1. What is the weather like generally for each commute?
  2. How long are the commutes?
  3. How close do you both live to a train station?
  4. Which country do you live in?

If you live in Oslo next to a train station and your office is also next to a train station and the distance between the two is 50-miles, then the answer is clear.

If you colleague lives in Holland and is 10-miles from the nearest train station, and the commute is 10-miles on a bike, then the answer is clear.

Why can't the same principle apply to workloads?

If you take a look at Oracle as a Service on a private cloud as an example, then the answer is a train, and let me explain why....

  • Oracle does not play nice with VMWare... the default position of Oracle is don't virtualise our database unless you do it on our virtualisation software
  • Oracle typically charge for a license per every two cores of x86
  • A typical Oracle license will cost £25K over 3 years

So if you have 288 cores of x86 for instance then you need 144 licenses or put another way you need to give the Larry Ellison yacht Fund £3.6m over 3-years.

Now imagine a weird and wonderful black magic platform existed where you could handle the same workload on a 12 core machine, which need 12 Oracle licenses, which had no problems with its virtualisation layer supporting Oracle.  The math goes someithing like this:

12 cores = 12 licenses = 12 x £25,000 = £300,000

Or put another way Larry has to wait another year to buy that racing yacht.

Or to go back to my original analogy still want to ride to work?

No comments:

Post a Comment