InfiniBand interconnects---The road to PetaFLOP clusters

Abstract

How far are we from achieving the next level of scalability in clustered supercomputing: the PetaFLOP machine? Is the end in sight? With popular speculation that 2008 will be the year, what are the important considerations for designing a PetaFLOP system? Can today's technologies support PetaFLOP clusters?A PetaFLOP machine - with a top speed of 1 quadrillion mathematical computations per second - will require thousands upon thousands of multicore CPUs, higher CPU speeds, and enormous amounts of disk space. These technologies will require massive amounts of I/O to avoid being the bottleneck.This session will look at the issues associated with building a PetaFLOP cluster and propose a solution using InfiniBand-based technologies that enable high performance interprocessor communication, high speed storage (through native InfiniBand storage, scalable file I/O), scalability (through high port count switches) and management. Sample configurations will be provided.