Date of Award
Summer 8-15-2021
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Parallel systems are pervasive nowadays. Specifically, modern computers have embraced multicore architectures due to the difficulties of exploiting higher clock speeds on single-core CPUs. However, parallel programming is challenging. Determinacy race, in particular, is a common pitfall when writing task-parallel code. It can easily lead to non-deterministic behavior of the parallel program and therefore a determinacy race is often considered as a bug. Unfortunately, such bugs are hard to debug because they do not necessarily produce obvious failures in every single execution.
To ease the debugging process of determinacy races in task-parallel code, this dissertation proposes several provably and practically efficient parallel race detection algorithms. Unlike prior works mostly target fork-join parallelism, we focus on less structured but important programming paradigms – pipeline parallelism and futures. In addition, we build an efficient runtime system for scheduling futures, which is not only a facility to study the race detection problem for futures but also useful in practice. Finally, this dissertation investigates mechanisms that optimize the access history of race detectors, which provides significant additional boost to the performance.
Language
English (en)
Chair
I-Ting Angelina Lee
Committee Members
Kunal Agrawal