Analysis of Meta-level Communication for Distributed Resource Allocation Problems

Matthew Saponaro, Keith Decker

Abstract: We examine the effects of meta-level-communication in the DSRAP (Distributed Sequential Resource Allocation Problem). In DSRAP, independent tasks are categorized into different types, where each task belonging to a particular task type shares a known distribution of task arrivals, durations, reward rates, maximum waiting times, and resource demands. We first look at a single task type DSRAP (SDSRAP) and develop an analytical model of the effect of meta-level communication about load on system performance for first-in-first-out (FIFO) local scheduling agents that forward tasks based on load. Through our analytical models and empirical results, we show how the frequency of communication affects performance for SDSRAP problems with one resource and task type. We then quantitatively measure the impact of meta-level communication on system performance with respect to the global measures of the system's load balance. We validate our analytical model's predictions experimentally, showing, e.g., as system load becomes unbalanced, performance decreases; organizational structure significantly impacts agent performance; and agents that can communicate and distribute tasks to neighbors significantly outperform agents working individually. Through our analysis on FIFO, routing algorithms, and our policy agents, we provide a framework for analyzing more complex task schedulers and task routing algorithms.