The HPC Stack

I was reading the news today and came across this article. The author believse that there should be no seperation between cluster level message passing and hardware acceleration. Furthermore, he believes that application writers should only have to use one API to an entire cluster with specialized hardware. I would think that the best way to accomplish this is to use a hybrid approach where a specific API is used for cluster messaging and another API is used for hardware acceleration. This seperation is mostly beneficial from a project management standpoint as it can reduce the scope of the message passing interface and the hardware accelerator considerably and allows the projects to be run completely independent. Yes, a programmer would have to learn two API's instead of one, but accelerating a program using hardware would have a different API from message passing no matter the abstraction. Therefore, adding an API on top of this would still have almost the same number of calls as the number of calls in each library added together.