Multi-Core Applications

Multicore technology is forcing concurrency into the mainstream of programming. In general, the application developer will have little control over the execution environment, such as the number of cores. New tools will be needed to help application developers identify parallelism, write correct and scalable code, and analyze performance bottlenecks. We propose to use dynamic instrumentation to build a general framework for implementing such tools for the Cilk++ multithreaded programming environment. Specifically, we propose to build a build a race-detection engine and a performance-analysis engine that can analyze work, span, parallelism, granularity, lock contention, and false sharing. We also propose to explore the feasibility of automatic performance tuners and parallelism estimators within this framework. A key property of our proposed tools is that they operate on the logical dependencies among the threads of a multithreaded computation, rather than on how the computation happens to be scheduled in any specific execution run.