[pl-seminar] "Improve Concurrent Software reliability via an effect-oriented approach", Zhang

[Aditya Thakur <adi@xxxxxxxxxxx>]

Title: Improve Concurrent Software reliability via an effect-oriented approach

Speaker: Wei Zhang

When: 4 pm, Fri. Dec 21st, 2012
Where: 3331CS

(Snacks will be served.)

Abstract:

Concurrent software reliability is becoming a more and more critical issue
these days. It is important to detect the concurrency bugs before the
software release and to recover software from the failure if concurrency
bugs slip into production runs.

In the past, much concurrency bug detection research has focused on the
cause of concurrency bugs(i.e. data race), which lead to both false
positive problems (as many data races are benign) and false negative
issues ( as bug detector designed for one type of bugs will miss other
types). For failure recovery, much research work has taken the whole
program rollback and re-execute approach, which leads to degraded
performance.

My work starts from the effect side of the concurrency bugs and exploits
the previously little know characteristics of effect of concurrency
bugs(i.e. error prorogation and failure type) to design a framework to
detect concurrency bug and recover concurrent software from failures. The
result shows that for the set of benchmarks(MySQL, Apache, Mozilla,
SPLASH2 etc) we evaluated, the bug detection tool achieves about 10X fewer
false positives while being able to detect unknown bugs; and the failure
recovery tool can survive all the failures without any help from bug
detection tool while incurring at most 0.2% overhead.

The topics of the talk are based on my work of  effect-oriented
concurrency bug detection -- ConMem(ASPLOS'10) and ConSeq(ASPLOS'11); and
effect-oriented concurrency failure recovery -- ConAir(will appeared  in
ASPLOS'13).

-- 
http://cs.wisc.edu/~adi