To distribute or not to distribute, that is the question in petascale and beyond

ABSTRACT

While there is general consensus that computing platforms underlying the grid infrastructures will continue to evolve, variance in the speed of technology acceleration in HPC is causing many of the assumptions made in the early days of grid to no longer hold. Such divergence in the metrics, as well as wider proliferation of related technologies such as Web2.0 and Clouds, will be changing the optimal design of the overall grid infrastructure towards more centralization of resources, much like the modern-day Internet with its two tier-structure of clients versus the datacenters. Based on our recent experiences with our TSUBAME supercomputer, which is currently Japan's fastest machine according to the Top500, and its next petascale generation design thereof, we will discuss the future design of multi-petascale grids with such machines being constituent massive resource nodes instead of vast distribution, as well as more advanced notions of computing where distribution of compute resources might be fundamental.