“If Cloud HPC works for your application(s), it is almost always going to be cheaper and easier than buying a new cluster.”

I have been with many companies as both HPC admin and HPC subject matter expert and the data sizes would really be a stumbling block to effectively using HPC clusters in the cloud. The network bandwidth that would be needed to allow users to get timely results in and out of a remote cluster would in many cases be prohibitively expensive.

Cloud HPC might be a suitable alternative if these batches of runs are infrequent and the cloud technical support available matches the user’s needs (they often don’t).

If frequent access to HPC systems over several years is required, then that Cloud money might be better spent, possibly contributing to local infrastructure that is managed by someone else.

The data security issues and data transfer times associated with using an HPC Cloud can also be show stoppers and always need to be considered.

what bugs me about the current cloud noise is that a cloud is just a cluster with VM-enforced isolation. you still need a scheduler (perhaps two levels), and pay a noticeable cost in performance. and now you get to maintain your own OS+software stack. but at least you’ll never run ‘ps’ and see someone you don’t know ;)