Speaker

John Shalf (Lawrence Berkeley National Laboratory)


Title

Green Flash: Exascale Computing on a Petascale Power Budget


Abstract

The past few years has seen a sea change in computer architecture that will impact every facet of our society as every electronic device from cell phone to supercomputer will need to confront parallelism of unprecedented scale. Energy has become the leading design constraint for future computing hardware, that has lead to a stall in the past decade of relentless performance improvements for HPC systems. As we look towards the next decade of leading edge HPC system architectures, the need to switch to a geometric growth path in system concurrency is leading to reconsideration of interconnect design, memory balance, and I/O system design that will have dramatic consequences for the design of future HPC applications and algorithms. The required reengineering of existing application codes will likely be as dramatic as the migration from vector HPC systems to Massively Parallel Processors (MPPs) that occurred in the early 90fs. Such comprehensive code reengineering took nearly a decade, so there are serious concerns about undertaking yet another major transition in our software infrastructure.

This presentation explores the fundamental device constraints that have led to the recent stall in CPU clock frequencies, and the consequences for future system designs. It then examines how multi-targeted auto-tuning and automated code generation systems can address the challenges of optimizing codes for increasingly complex multicore (or manycore) hardware. We then examine the power-efficiency benefits of tailoring computer designs to the problem requirements. Finally, we show how tightly integrated hardware/software co-design methods, that employ auto-tuning methodologies, can achieve power efficiency and performance improvements hundreds of times better than following conventional industry trends. Our approach is elucidated by a design study of an ultra-efficient computing system optimized for climate modeling called Green Flash.