“Rethinking Quantitative Approaches to Computer System Design in the Power Age”
Thursday, Oct. 3, 1:00pm
310 Larsen Hall
From instruction pipelining to multi-issue and prefetching on hyper-threaded, multi-core processors, parallelism (or overlapped work) increases the performance of modern applications and systems. For more than two decades, the field has relied upon a quantitative approach to computer system design that leverages experimental measurements coupled with performance evaluation and prediction. However, despite the evolution, emergence, intricacy, and pervasiveness of designs that increase parallel overlap, application and system performance evaluation techniques have changed little since the 1990s.
A hypothesis of this work is that the elevation of power consumption to a first-class design constraint renders traditional quantitative approaches to system design obsolete. The key observation is that power management alters memory and processor throughput dynamically and that when combined with thread-level parallelism, significant performance overlap effects go undetected using best available parallel performance models. In this talk, I will show how emergent power management techniques require new insights to computation and memory overlap. To this end, I will discuss the Compute-Overlap-Stall (COS) model of parallel computation that enables accurate prediction of the simultaneous performance impact of processor, memory, and thread throttling. The implication of our findings is that as power management techniques pervade, new quantitative approaches that isolate the effects of overlap are essential to evaluating future computer system designs.
Dr. Kirk W. Cameron is Professor of Computer Science, Director of the stack@cs Center for Computer Systems, and Associate Department Head for Research and Engagement at the Department of Computer Science at Virginia Tech. He is a Green HPC pioneer (Green500, SPECPower, PowerPack, grano.la) and his software has been downloaded by more than 500,000 people in 160+ countries. In addition to NSF and DOE Career Awards, IBM and AMD Faculty Awards, over 100 publications and the HPDC 2017 Best Paper Award, he conceived and created SeeMore, a 256-node kinetic sculpture of Raspberry Pi’s to introduce parallel computing to kids aged 2 to 92. His work consistently appears in The New York Times, The Guardian, Time, Newsweek, etc. and SeeMore was named the second best RPi project of all time by MagPi Magazine. Prior to completing his Ph.D. at LSU in 2000, he graduated with a B.S. in Mathematics from UF in 1995 and still bleeds orange and blue.