Sudip Dosanjh: Select Publications

Extreme data science is becoming increasingly important at the U.S. Department of Energy's National Energy Research Scientific Computing Center (NERSC). Many petabytes of data are transferred from experimental facilities to NERSC each year. Applications of importance include high-energy physics, materials science, genomics, and climate modeling, with an increasing emphasis on large-scale simulations and data analysis. In response to the emerging data-intensive workloads of its users, NERSC made a number of critical design choices to enhance the usability of its pre-exascale supercomputer, Cori, which is scheduled to be delivered in 2016. These data enhancements include a data partition, a layer of NVRAM for accelerating I/O, user defined images and a customizable gateway for accelerating connections to remote experimental facilities.

High Performance Computing architectures are expected to change dramatically in the next decade as power and cooling constraints limit increases in microprocessor clock speeds. Consequently computer companies are dramatically increasing on-chip parallelism to improve performance. The traditional doubling of clock speeds every 18-24 months is being replaced by a doubling of cores or other parallelism mechanisms. During the next decade the amount of parallelism on a single microprocessor will rival the number of nodes in early massively parallel supercomputers that were built in the 1980s. Applications and algorithms will need to change and adapt as node architectures evolve. In particular, they will need to manage locality to achieve performance. A key element of the strategy as we move forward is the co-design of applications, architectures and programming environments. There is an unprecedented opportunity for application and algorithm developers to influence the direction of future architectures so that they meet DOE mission needs. This article will describe the technology challenges on the road to exascale, their underlying causes, and their effect on the future of HPC system design.