We study the process of multi-agent reinforcement learning in the context of load balancing in a distributed system, without use of either central coordination or explicit communication. We first define a precise framework in which to study adaptive load balancing, important features of which are its stochastic nature and the purely local information available to individual agents. Given this framework, we show illuminating results on the interplay between basic adaptive behavior parameters and their effect on system efficiency. We then investigate the properties of adaptive load balancing in heterogeneous populations, and address the issue of exploration vs. exploitation in that context. Finally, we show that naive use of communication may not improve, and might even harm system efficiency. 1. Introduction This article investigates multi-agent reinforcement learning in the context of a concrete problem of undisputed importance -- load balancing. Real life provides us with many exampl...

In the past, several process migration facilities for distributed systems have been developed. Due to the complex nature of the subject, all those facilities have limitations that make them usable for only limited classes of applications and environments. We discuss some of the usual limitations and possible solutions. Specifically, we focus on migration of groups of collaborating processes between Unix systems without kernel modifications, and from this we derive the design for a migration system. First experiences with our implementation show that we reach performance figures for the migration that are close to those of real distributed operating system.

Load-balancing systems use workload indices to dynamically schedule jobs. We present a novel method of automatically learning such indices. Our approach uses comparator neural networks, one per site, which learn to predict the relative speedup of an incoming job using only the resourceutilization patterns observed prior to the job's arrival. Our load indices combine information from the key resources of contention: CPU, disk, network, and memory. Our learning algorithm overcomes the lack of job-specific information by learning to compare the relative speedups of different sites with respect to the same job, rather than attempting to predict absolute speedups. We present conditions under which such learning is viable. Our results show that indices learnt using comparator networks correctly pick the best destination in most cases when incoming jobs are short; accuracy degrades as execution time increases. 1. INTRODUCTION This papers addresses the computation of workload measures in dis...

This paper describes Dynamic Workload Generator (DWG), a facility for generating realistic and reproducible synthetic workloads for use in load-balancing experiments. For such experiments, the generated workload must not only mimic the highly dynamic resource-utilization patterns found on today's distributed systems but also behave as a real workload does when test jobs are run concurrently with it. The latter requirement is important in testing alternative load-balancing strategies, a process that requires running the same job multiple times, each time at a different site but under an identical network-wide workload. Parts of DWG are implemented inside the operating-system kernel and have complete control over the utilization levels of four key resources: CPU, memory, disk, and network. Besides accurately replaying network-wide load patterns recorded earlier, DWG gives up a fraction of its resources each time a new job arrives and reclaims these resources upon job completion. The latt...

This report presents a load sharing heuristic for distributed computing on workstation clusters. The approach is novel in that it combines the use of predicted resource requirements of processes (CPU-time, memory requirements, density of the I/O-stream) and a fuzzy logic controller which makes the placement decision. The heuristic is distributed, i.e. each node runs a copy of the prediction and load sharing code, and its implementation is based on PVM. Using a benchmark program (Choleski factorization) experiments were conducted to compare the proposed heuristic against standard PVM and an older version of the presented heuristic without the fuzzy logic controller. Department of Computer Science, University of Rostock, Rostock, D-18051 Rostock, Germany; Phone: +49 381 498 3403, Fax: +49 381 498 3366, E-mail: hunger@informatik.uni-rostock.de y Department of Mathematics, Technical University Ilmenau, D-98684 Ilmenau, PF 10 0565, Germany, Phone: +49 3677 69 3630, Fax: +49 3677 69 3206...

A number of load sharing/balancing mechanisms was developed in order to improve the execution of a distributed application in a workstation cluster. The contribution presents an adaptive load sharing system basing on a load prediction for the computing nodes of the system. Using an online calculation of stochastic parameters of processes the quality of load sharing could be improved once more. The reachable results will be predicted using a Petri Net simulation. Furthermore the execution of a benchmark program in the new environment will be discussed. 1. INTRODUCTION Distributed programming on workstation clusters became more and more important to solve a number of problems faster and more effectively. Obviously, the machines of such an architecture are not exclusively used by any user. Therefore a number of load sharing and load balancing mechanisms were developed to optimize the execution of processes on a workstation cluster architecture. (for an overview/classification see (Baker,...

This paper studies load balancing issues for classes of problems with certain bisection properties. A class of problems has ff-bisectors if every problem in the class can be subdivided into two subproblems whose weight is not smaller than an ff-fraction of the original problem. It is shown that the maximum weight of a subproblem produced by Algorithm HF, which partitions a given problem into N subproblems by always subdividing the problem with maximum weight, is at most a factor of b1=ffc \Delta (1 \Gamma ff) b1=ffc\Gamma2 greater than the theoretical optimum (uniform partition). This bound is proved to be tight. Two strategies to use Algorithm HF for load balancing distributed hierarchical finite element simulations are presented. For this purpose, a certain class of weighted binary trees representing the load of such applications is shown to have 1=4-bisectors. This establishes a performance guarantee of 9=4 for load balancing in this case. 1 Introduction Load balancing is one of...

Contents 1 Introduction 9 1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2.1 The original idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2.2 The PRP system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.3 Chapter division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.4 To future readers of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2 Problem definition 13 2.1 The concept of Parallel Recursive Procedures . . . . . . . . . . . . . . . . . . 13 2.1.1 An example introducing the principal ideas . . . . . . . . . . . . . . . 13 2.1.2 Fanout and the administration of the servers . . . . . . . . . . . . . . 14 2.1.3 Breadth first versus depth first . . . . . . . . . . . . . . . . . . . . . . 16 2.1.4 Returning from the recursive proced

Effective load-balancing policies use dynamic resource information to schedule tasks in a distributed computer system. In this paper, we present a novel method for automatically learning such policies. At each site in our system, we use a comparator neural network to predict the relative speedup of an incoming task using only the resource-utilization patterns obtained prior to the task's arrival. Outputs of these comparator networks are broadcast periodically over the distributed system, and the resource schedulers at each site use these values to determine the best site for executing an incoming task. The delays incurred in propagating workload information and tasks from one site to another, as well as the dynamic and unpredictable nature of workloads in multiprogrammed multiprocessors, may cause the workload pattern at the time of execution to differ from patterns prevailing at the times of load-index computation and decision making. Our loadbalancing policy accommodates this uncerta...

Abstract. Two new performance indices (PIV- Performance Index Vector and WPIV – Weighted Performance Index Vector) are presented in this article. Those indices to evaluate heterogeneous computing systems are based on a Euclidian metric. Aiming to maximize the use of the machines, the proposed indices are a combination of several usual indices and the results of their evaluation through a simulator show an appropriate behavior for different kinds of applications.