Building a Foundation for the Next PVM: Petascale Virtual Machines

Abstract

High performance computing is on the verge of another major jump in scale and now is the time for us to begin creating superscalable applications and systems software if we are to exploit the coming machines. In the early 1990s there was a major jump in scalability from a few hundred to a few thousand processors. Throughout the 1990s the size of the largest machines remained around this scale and PVM and MPI were key to making these machines useable.

Now we are seeing changes such as SETI@home and IBM’s Blue Gene systems that suddenly jump the scale of machines to 100,000 processors or more. Our software environment and applications are unprepared for this jump. This talk looks at the common software needs between these very different ways to get to this scale. The talk will argue that software may need to fundamentally change to allow us to exploit machines that approach a petascale. Interspersed with this discussion will be results from research going on today to increase the scalability, fault tolerance, and adaptability of MPI and systems software.