Tolerating asymmetric data races with minimal hardware support

Data races are a common type of concurrency bug in parallel programs. An important type of race that has not received much attention is Asymmetric data races. In these races, the state of well tested, correct threads is corrupted by racing threads from external, typically third-party code. Current schemes to detect and tolerate these races are software based, and have substantial execution overhead. This thesis proposes the first scheme to detect and tolerate asymmetric data races in hardware. The approach, called Pacman, induces negligible execution overhead and requires minimal hardware modifications. In addition, compared to past software-based schemes, Pacman eliminates deadlock cases. Pacman is based on using hardware address signatures to detect the asymmetric races. Processor, cache coherence, and protocol messages remain unchanged. We evaluate Pacman for all the SPLASH2 and PARSEC applications. Our results show that Pacman is effective and has minimal overhead. ii

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2010

Author dc:creator

• Qi, Shanxiang

Contributors dc:contributor

• Torrellas, Josep

• computer architecture
• software reliability
• concurrency bugs

Statement dc:rights

• Copyright 2010 Shanxiang Qi

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/15995

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/15995