CSR: Small: Collaborative Research: Pursuing High Performance on Clouds and Other Dynamically Heterogeneous Computing Platforms

Northeastern University, Boston MA

Investigators

Abstract

Emerging technologies have led to revolutionary computing platforms such as computing clouds and many genres of computing grids. These platforms promise to make high-performance computing platforms accessible to the public. Realizing this promise, though, requires one to cope with the platforms' dynamic heterogeneity, i.e., the fact that their constituent computers' relative powers and speeds can change at unpredictable times and in unpredictable ways; for instance, shared computers may slow down or speed up significantly because of unpredictable changes in workloads. The research of the PIs is developing a transformative computing paradigm that will enable high-performance computing on these platforms. The new paradigm replaces traditional schedulers' attempts to accommodate the particulars of a computing platform---a goal that dynamic heterogeneity confutes---by orchestrating a complex computation in a way that honors the relevant details of the computation's inherent structure. In this way, the paradigm increases opportunities for executing independent tasks in parallel (i.e., simultaneously), thereby completing computations faster. Preliminary assessments---via simulated competitions with common computation schedulers---suggest that the new paradigm often completes computations faster than competing strategies by double-digit percentages. The challenge is to realize the paradigm in a computationally efficient manner, and the PIs are pursuing a variety of avenues toward achieving such efficiency. Because of the complexity of dynamically heterogeneous platforms, this goal requires advances in the technology of both scheduling algorithms and simulation software. The PIs are uniquely qualified to produce such advances because of their complementary expertise in algorithms, applications, and systems.