Semi-empirical Multiprocessor Performance Predictions1

architecture behavior, most analytical methods use hierarchical models. In [3], Adve provides a framework for parallel program performance prediction models which well characterizes most of the existing models by a hierarchy of higher and lower level models. In the higher-level component, task graphs [11, 13] are usually used to represent the task-level behavior of the program. A task graph is a directed acyclic graph in which each vertex represents a task and each edge represents the precedence relationship between a pair of tasks. There is no internal parallelism inside a task, and a task must be executed sequentially. This higher-level model component computes the overall execution time assuming individual task execution times are known. The lower-level components represent systemlevel effects and are usually simulated by stochastic processes [6, 13] or by some type of system overhead function [10]. Individual task execution times are computed from lower-level components. Thomasian and Bay [13] propose a two-level model for a class of programs which can be represented by directed acyclic graphs. At the higher level, the system behavior is specified by a Markov chain whose states correspond to the combination of tasks in execution. At the lower level, the transition rates among the states of the Markov chain are computed using a queueing network solver, which determines the throughput of the computer system for each system state. Vrsalovic et al. [15] develop an analytic model for predicting the performance of iterative algorithms. Using the same approach, we predict the execution performance of a program with a larger number of iterations based on the performance of the same program with a small number of iterations. However, the method in [13] focuses on iterative algorithms and models the decomposition of a program into processes by using pure analytic functions. The model proposed in this paper focuses on broader categories of applications. The hierarchical model in this paper distinguishes deterministic factors from non-deterministic performance factors, and implicit communications from explicit communications. Both analytic and experimental methods are combined performance prediction. Kapelnikov et al. [6] propose a methodology that embodies two modeling domains: the program domain and the physical domain. In the program domain, a graphical model JOURNAL OF PARALLEL AND DISTRIBUTED COMPUTING 39, 14–28 (1996) ARTICLE NO. 0151