Cilk: efficient multithreaded computing

This thesis describes Cilk, a parallel multithreaded language for programming contemporary shared memory multiprocessors (SMP's). Cilk is a simple extension of C which provides constructs for parallel control and synchronization. Cilk imposes very low overheads | the typical cost of spawning a parallel thread is only between 2 and 6 times the cost of a C function call on a variety of contemporary machines. Many Cilk programs run on one processor with virtually no degradation compared to equivalent C programs. We present the \workrst" principle which guided the design of Cilk's scheduler and two consequences of this principle, a novel \two-clone" compilation strategy and a Dijkstra-like mutual-exclusion protocol for implementing the ready queue in the work-stealing scheduler. To facilitate debugging of Cilk programs, Cilk provides a tool called the Nondeterminator-2 which nds nondeterministic bugs called \data races". We present two algorithms, All-Sets and Brelly, used by the Nondeterminator-2 for nding data races. The All-Sets algorithm is exact but can sometimes have poor performance; theBrelly algorithm, by imposing a locking discipline on the programmer, is guaranteed to run in nearly linear time. For a program that runs serially in time T , accesses V shared memory locations, and holds at most k locks simultaneously, Brelly runs in O(kT (V; V )) time and O(kV ) space, where is Tarjan's functional inverse of Ackermann's function. Cilk can be run on clusters of SMP's as well. We de ne a novel weak memory model called \dag consistency" which provides a natural consistency model for use with multithreaded languages like Cilk. We provide evidence that Backer, the protocol that implements dag consistency, is both empirically and theoretically e cient. In particular, we prove bounds on running time and communication for a Cilk program executed on top of Backer, including all costs associated with Backer itself. We believe this proof is the rst of its kind in this regard. Finally, we present the MultiBacker protocol for clusters of SMP's which extends Backer to take advantage of hardware support for shared memory within an SMP. Thesis Supervisor: Charles E. Leiserson Title: Professor of Computer Science and Engineering