SCHEDULING TREE DAGS ON PARALLEL ARCHITECTURES

Citation
K. Kalpakis et Y. Yesha, SCHEDULING TREE DAGS ON PARALLEL ARCHITECTURES, Algorithmica, 15(4), 1996, pp. 373-396
Citations number
13
Categorie Soggetti
Computer Sciences",Mathematics,Mathematics,"Computer Science Software Graphycs Programming
Journal title
ISSN journal
01784617
Volume
15
Issue
4
Year of publication
1996
Pages
373 - 396
Database
ISI
SICI code
0178-4617(1996)15:4<373:STDOPA>2.0.ZU;2-O
Abstract
We provide optimal within a constant explicit upper bounds on the make span of schedules for tree-structured programs on mesh arrays of proce ssors, and provide polynomial-time algorithms to find schedules with m akespan matching these bounds. In particular, we show how to find, in polynomial time, a (nonpreemptive) schedule for a binary tree dag with n unit execution time tasks and height h on a d-dimensional mesh arra y with m processors and links of unit bandwidth and unit propagation d elay whose makespan is O(n/m + n(l/(d+1)) + h), i.e., optimal within a constant factor. Further, we extend these schedules to bounded degree forest dags with arbitrary positive integer execution time tasks and to meshes when the propagation delay of all the links is an arbitrary positive integer. Thus, we provide a polynomial-time approximation alg orithm for an NP-hard problem, with a performance ratio that is a cons tant. We also show how to schedule tree dags on any parallel architect ure that satisfies certain natural, not very restrictive, conditions t hat are satisfied by most parallel architectures used in practice. Let epsilon be a fixed positive real number. We provide polynomial time c omputable schedules for binary tree dags with n unit execution time ta sks and height h is not an element of (g(n)n(-epsilon), g(n) log n) on any parallel architecture satisfying those conditions, with unit band width and unit propagation delay Links, with optimal up to a constant makespan O(g(n) + h), where g is a function that depends only on that architecture. The number of processors used is optimal within a consta nt factor if h less than or equal to g(n)n(-epsilon), and is optimal w ithin an O(log n) factor if h greater than or equal to g(n) log n. As an example, for hypercube and complete binary tree architectures, we a chieve optimal within a constant makespan O(h) when h = Omega(log(2) n ), using an optimal within an O (log n) factor number of processors. F urther, we extend these schedules to the case of bounded-degree forest dags with tasks of arbitrary positive integer execution times and arc hitectures when the propagation delay of all the links is a given arbi trary positive integer.