BOUNDS ON THE SCALABILITY OF BAG-OF-TASKS APPLICATIONS RUNNING ON MASTER-SLAVE PLATFORMS

Abstract

Bag-of-Tasks applications are parallel applications composed of independent (i.e., embarrassingly parallel) tasks that do not communicate with each other, may depend upon one or more input files, and can be executed in any order. Each file may be input for more than one task. A common framework to execute BoT applications is the master-slave topology, in which the user machine is used to control the execution of tasks. In this scenario, a large number of concurrent tasks competing for resources (e.g., CPU and communication links) severely limits the scalability. In this paper we studied the scalability of BoT applications running on multi-node systems (e.g. clusters and grids) organized as master-slave platforms, considering two communications paradigms: multiplexed connections and efficient broadcast. We prove that the lowest bound possible on the isoefficiency function for master-slave platforms is achievable by those platforms that have an O(1) efficient broadcast primitive available. We also analyze the impact of output file contention in scalability, under different assumptions. Our study employs a set of simulation experiments that confirms and extends the theoretical results (e.g. by simulating TCP links).