Abstract

In this paper, the problem of efficient scheduling k-strict multithreaded computations is studied. A k-strict multithreaded computation is one in which all data dependencies of a thread go to ancestors of the thread, that they are located in at most k higher levels, than that of the thread, in the activation tree. Two scheduling algorithms that solves efficiently this problem are presented here. Both of them use the work-stealing technique, in which a processor that doesn't have work to do, steals threads from other processors. Our results achieve linear speedup and linear expansion of memory, for constant k (that is k 2 O(1)). However, they are general and hold for all values of k. Thus, we nontrivially extend here the results of Blumofe and Leiserson [1], which examined the case k = 1. Keywords: multithreaded computation, scheduling, parallel computation, load balancing. 1 Introduction Dynamically growing "multithreaded" computations are quite common for parallel computers. To exec...

Citations