Concurrency bugs are caused by unsynchronized interaction among threads. This complicated cause makes concurrency bugs much more difficult to detect and recover from than sequential bugs. With multi-threaded software becoming prevalent in the multi-core era, concurrency bugs are widespread. They frequently escape in-house bug detection and severely hurt system reliability during production runs.

This talk will present an effect-oriented approach to in-house concurrency-bug detection and production-run concurrency-bug failure recovery. I will first use a characteristics study to demonstrate that concurrency and sequential bugs have drastically different causes but mostly similar effects. I will then present two bug detection tools, ConMem and ConSeq. Guided by the effect patterns of concurrency bugs, ConMem and ConSeq can detect concurrency bugs before they manifest with higher coverage and accuracy than traditional cause-oriented approaches. Of course, no bug detection tool is perfect. Next, I will discuss how we build ConAir to handle bugs that escape into production runs. ConAir is a static code transformation tool that enables automatic failure recovery through rollback reexecution. Different from traditional rollback-reexecution techniques, ConAir leverages the effect patterns of concurrency bugs to avoid memory checkpoint and multi-threaded rollback. As a result, it incurs negligible run-time overhead (<1%) with no OS/hardware change. I will conclude the talk by discussing other research in my group that tackles concurrency bugs and performance bugs.