HPDC is the premier computer science conference for presenting new research related to high-performance parallel and distributed systems used in both science and industry. Of approximately 100 submissions, only 19 papers were accepted, and just one received the Best Paper Award.
The winning paper presents the Compute-Overlap-Stall (COS) Model of parallel performance, a first-of-a-kind framework for accurately capturing the simultaneous, combined effects of three operating modes—dynamic voltage and frequency scaling (DVFS), dynamic memory voltage and frequency throttling (DMT), and dynamic concurrency throttling (DCT). COS is expected to be key in helping future high-performance systems meet the challenging demands of parallel scientific applications.
“This is an exciting new model that sheds light on the combined effects of variations in memory speed with techniques such as DVFS and concurrency throttling,” says León.
Future high-performance, scientific applications will have a high degree of parallelism and run in environments of enormous scale but limited power. Efficiency will be key to achieving the promise of exascale computing. Emerging systems will have large numbers of configurable operating modes that provide unprecedented control of processor speed and memory frequency and bandwidth. Unfortunately, little is known about the combined effects of these operating modes and thread concurrency on performance and execution time.
Figure: The key insight to the Compute-Overlap-Stall (COS) Model is the separation of memory and compute overlap from pure compute and pure memory stalls. (Click to enlarge.)
“Modeling the combined effects of DVFS, DMT, and DCT is incredibly challenging,” says León. “Capturing the interactive performance effects of a highly configurable problem space could be intractable in parallel, high-performance environments. Furthermore, the interactive effects of these modes are likely to be non-linear, complicating efforts to identify simple, but useful analytical models of performance.”
According to León, COS is based on a simple observation. Past models of performance tend to combine the overlap of compute and memory performance into either compute time or memory stall time. However, the overlap is so complex when these operating modes change, that it must be modeled independently. The COS Model allows each term to be modeled independently of the others.
“The COS Model may just be the breakthrough that transforms our understanding of computational performance,” says León. “I’m thrilled the conference recognized the potential of this work.”
High performance computing plays a crucial role in enabling the missions of the Laboratory. While exascale computing is expected to allow unprecedented levels of science, it poses great challenges in the areas of parallelism, reliability, and energy efficiency. León’s research could benefit the mission focus areas at LLNL by enabling larger problem sizes, greater simulation accuracy, or more detailed simulations.