Cloud-onomics for HPC

04 Feb Cloud-onomics for HPC

Let’s suppose you need to do some HPC. You have $50K to buy a cluster or use an HPC Cloud. For the sake of the example, let’s also assume your application works equally well with a cloud or local cluster. What should you do?

The cost of an HPC Amazon EC2 instance is $1.60 and hour. That will give you a Cluster Compute Quadruple Extra Large Instance with 23 GB of memory, 33.5 EC2 Compute Units (2 x Intel Xeon X5570, quad-core “Nehalem” architecture). Note, cheaper instances are available, but with virtual cores. Let’s assume you need eight of these for a total of 64 cores.

The amount of compute time you could buy with your budget is {$50,000/(8 x $1.60)} approximately 3900 hours (or 162 days). Remember that this is 24×7 computing. So lets spread this over two years. Thus if you use the Amazon Cloud you will exhaust your budget in two years. If you buy a cluster, you will have a resource beyond two years, but is that a better deal?

It all depends. Some might argue that after two years the initial cluster you purchased is “old,” but that is not really the case. Individual cores have not had a significant increased in clock speed over the last 3-4 years and don’t expect any great leaps in the future. You can expect to see more cores per socket, however. Therefore, your cluster is going to be able to useful work well beyond the two year limit.

If you believe an in-house resource would be more valuable (i.e. you expect your HPC workload to continue for four years or more), then buying a cluster may make more sense. If you have have a project that will be complete after two years, the Cloud approach may be a good idea. Indeed, assume further that you need to have a large amount of computing for the initial six months (maybe you need 16 EC2 instances) and then your use will get less as the project continues (for the remaining 18 months). In this case, a Cloud makes good sense.

There is also a middle ground between the Cloud and resident cluster. Consider spending spending $25K of your budget on a small in-house cluster to cover the base level of work coupled with $25K for Cloud HPC to support the heavy work periods. This is technique has been called “Cloud Bursting” and I believe it will become the predominant method of using the HPC Cloud. We will always need clusters on the “ground.”