CAREER: Deterministic Shared Memory Multiprocessing: Vision, Architecture, and Impact on Programmability

University Of Washington, Seattle WA

Investigators

Abstract

Software innovation typically relies on performance improvements of the underlying hardware. However, technology limitations hinder further significant progress in single-thread performance. Therefore, the software industry has the immense problem of rethinking its software development process and techniques to adopt multicore systems. Popularizing parallel programming is a Grand Research Challenge for the systems community [CRA]. Being able to leverage the full potential of multicores would put us back into exponential growth of usable performance as well as lead to significant power savings. One of the main reasons why parallel programming is hard is that parallel code in current multicore systems can execute nondeterministically. Each time a multicore executes a parallel application, it can produce a different output even if supplied with the same input. This frustrates debugging efforts and limits the ability to properly test parallel code, becoming a major obstacle to widespread adoption of parallel programming. This project poses broad intellectual questions with far-reaching implications in modern computer systems: Can nondeterminism be removed from shared-memory multiprocessor systems without degrading performance? What are the trade-offs in designing deterministic multiprocessor systems? What are the implications and uses of deterministic behavior in programmability? The PI plans to answer these questions by devising efficient, general purpose, fully deterministic shared memory multiprocessor systems and demonstrating that they can enable significant changes in how parallel programs are written, tested and deployed. An integral part of the concurrency challenge is education, so this project also aims to develop a graduate and undergraduate curriculum that will teach students about concurrency principles and practical parallel programming.