U.S. Department of Energy

Pacific Northwest National Laboratory

Commodity-Based Approaches to Scaling Irregular Computation

From DIC

Jump to: navigation, search
Today, the Cray XMT solves certain large “irregular problems” arising in today’s national security, engineering and science applications with performance unmatched by any other computer system. Going forward, these applications will demand ever increasing performance. To meet this demand requires further advances in computing technology. Unfortunately, redesigning the Cray XMT’s Threadstorm processor from the ground up would be prohibitively expensive. This evaluation task explores cost-effective alternatives.
We focus our evaluation of cost-effective alternatives to the Cray XMT on systems built largely or wholly from commodity hardware components. What we call “SoftXMT” are software-based system solutions leveraging Multicore Processors and Field Programmable Gate Arrays (FPGAs).
Multicore Processors are prevalent in today’s computers: from handhelds to warehouse-sized supercomputers, they represent the dominant paradigm in computing. Though the Cray XMT’s Threadstorm processors are slower on almost all computations than Multicore Processors, they do possess features that enable them to work together efficiently on large irregular problems. One of the key features is latency tolerance: Threadstorm processors keep themselves busy even while waiting for remote fetches of faraway data by juggling up to 128 subcomputations simultaneously. Doing so requires both an efficient mechanism for switching amongst these computational contexts -- the hands of the juggler -- and the capacity to keep track of the thousand or more references in flight -- the juggler’s eyes. Multicore Processors lack both hands and eyes. In performing our evaluation, we must determine whether or not software can compensate adequately for these missing features.
FPGAs are configurable components that can be programmed to deliver specific functionality at greater efficiency than multicore processors, though as general purpose computing devices they are inherently slower. By programming them to provide the hands and eyes -- as well as other Threadstorm features -- that Multicore Processors lack, we may be able to provide hardware compensation that puts commodity systems on par or ahead of the Cray XMT, even within its sweetspot for performance. Our evaluation will explore architectures incorporating FPGAs to see what it would take for them to provide a viable alternative.

Article Title: Commodity-Based Approaches to Scaling Irregular Computation

Article Added: 2010/09/18

Category(s): Hybrid Architectures

Last Update: 13 July 2011 | Pacific Northwest National Laboratory