Executing multithreaded programs efficiently

This thesis presents the theory, design, and implementation of Cilk (pronounced “silk”) and Cilk-NOW. Cilk is a C-based language and portable runtime system for programming and executing multithreaded parallel programs. Cilk-NOW is an implementation of the Cilk runtime system that transparently manages resources for parallel programs running on a network of workstations. Cilk is built around a provably efficient algorithm for scheduling the execution of fully strict multithreaded computations. Based on the technique of work stealing, this algorithm achieves time, space, and communication bounds that are all within a small constant factor of optimal. Using these performance guarantees, Cilk provides a simple abstraction of performance that allows programmers to predict accurately how program changes will affect execution time. Cilk-NOW provides high-level resource management services so that Cilk programs can run efficiently on a network of workstations. These services include adaptive parallelism and fault tolerance. With adaptive parallelism, programs execute on a pool of workstations that dynamically grows and shrinks in response to both the availability of “idle” machines and the availability of parallelism within the program. Cilk-NOW also provides transparent fault tolerance with a fully distributed checkpointing mechanism.