This paper describes AntSM, a system that uses the inherent parallelism of multi-threaded programs to reduce the overhead of statistical and invariant violations detection-based debugging tools. The run-time monitorin...
详细信息
ISBN:
(纸本)9783319099675;9783319099668
This paper describes AntSM, a system that uses the inherent parallelism of multi-threaded programs to reduce the overhead of statistical and invariant violations detection-based debugging tools. The run-time monitoring of these tools leads to high overheads. The key insight of the AntSM system is that this overhead can be reduced in parallelprograms by performing sampled monitoring across parallel regions of the program that are performing similar actions. AntSM implements this sampling using a combination of static and dynamic analyses to determine similar parts of the program executing in parallel and the number of threads executing those parts of the program. Experimental results, performed using the C-DIDUCE (a variant of DIDUCE for C) debugging tool on eleven Pthreads benchmarks from the PARSEC suite, show monitoring overhead is reduced by up to 18.14 times (and on average 8.73 times) on an eight-core machine relative to a naive port that performs no sampling.
Detecting data races is a hard problem when debugging shared memory parallelprograms, because the races could exhibit unpredictable results in execution of programs. It is even harder for programmers to decide whethe...
详细信息
ISBN:
(纸本)9783642355998
Detecting data races is a hard problem when debugging shared memory parallelprograms, because the races could exhibit unpredictable results in execution of programs. It is even harder for programmers to decide whether or how the reported data races can appear in the actual program executions. Previous techniques for detecting races cannot provide method to discover first races in parallelprograms with random synchronization. We present an algorithm, in this paper, which extracts first races to occur by comparing the concurrency of the traced information with the candidate accesses which are from a particular execution of parallelprograms.
暂无评论