Ten Reasons to Use Divisible Load Theory

Thomas G. Robertazzi State University of New York, Stony Brook T he increasing prevalence of multiprocessor systems and data-intensive computing has created a need for efficient scheduling of computing loads, especially parallel loads that are divisible among processors and links. During the past decade, divisible load theory has emerged as a powerful tool for modeling data-intensive computational problems. DLT originated from a desire to create intelligent sensor networks, but most recent applications involve parallel and distributed computing. The first published research on divisible load theory appeared in a 1988 doctoral dissertation by James Cheng— now director of venture technology at the Siemens Technology-to-Business Center in Berkeley, California—at the State University of New York, Stony Brook. Cheng originally sought to develop intelligent sensor networks that could make measurements, compute, and communicate, but his thesis committee regarded his early work as too theoretical and required more application-related material before giving their approval. Cheng’s original 1988 paper contains an intuitive proof of the DLT optimality principle, but a formal proof did not appear until five years later, following extensive search runs on an IBM mainframe in 1989. The theory’s linearity also gradually became apparent in the early 1990s. Since then, an international group of researchers from several countries has participated in a cooperative effort to develop DLT, generating more than 50 journal papers and two books (www.ece.sunysb.edu). Like other linear mathematical models such as Markovian queuing theory and electric resistive circuit theory, DLT offers easy computation, a schematic language, and equivalent network element modeling. Because DLT does not recognize precedence relations among data, it assumes that computation and communication loads can be partitioned arbitrarily among numerous processors and links, respectively. While it can incorporate stochastic features, the basic model does not make statistical assumptions, which can be the Achilles’ heel of a performance evaluation model.