We measure the distribution of lifetimes for UNIX processes and propose a functional form that fits this distribution well. We use this functional form to derive a policy for preemptive migration, and then use a trace-driven simulator to compare our proposed policy with other preemptive migration policies, and with a non-preemptive load balancing strategy. We find that, contrary to previous reports, the performance benefits of preemptive migration are significantly greater than those of non-preemptive migration, even when the memorytransfer cost is high. Using a model of migration costs representative of current systems, we find that preemptive migration reduces the mean delay (queueing and migration) by 35 -- 50%, compared to non-preemptive migration. 1 Introduction Most systems that perform load balancing use remote execution (i.e. non-preemptive migration) based on a priori knowledge of process behavior, often in the form of a list of process names eligible for migration. Althoug...

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...

) R. Luling, B. Monien, F. Ramme Department of Mathematics and Computer Science University of Paderborn, Germany e-mail : rl@uni-paderborn.de, bm@uni-paderborn.de, ram@uni-paderborn.de Abstract In this paper we compare six well known and two new load balancing strategies on torus and ring topologies of different sizes and workload characteristics. Through simulations on a large transputer network, we show that all strategies behave differently under the workload of process and data migration. The two new algorithms based on the gradient model method are shown to be robust to both kinds of workloads. Thus, these new algorithms are good candidates for distributed operating systems running on large networks, where the workload characteristics can not be determined in advance. 1 Introduction We study load balancing algorithms on large MIMD multiprocessor systems. The systems we consider are homogeneous and consist of autonomous processing elements (324 transputers in our case), which...

In a heterogeneous distributed real--time system, some nodes may experience more task arrivals than others, or tasks arrived at some nodes may have tighter laxities than those arrived at other nodes. In such an environment, transferring an unguaranteed task on a node to another node currently with the most abundant resources is not necessarily the best decision. We propose a new load sharing (LS) algorithm for real--time applications which takes into account the effect of future task arrivals on locating the best receiver for each unguaranteed task. Upon arrival of a task at a node, the node first checks whether or not it can complete the task in time using the minimum--laxity--first--served discipline. If the node cannot guarantee the arrived task or some of existing guarantees were to be invalidated as a result of inserting the task into its queue, then the node must locate a remote node to which each unguaranteed task will be transferred. The proposed LS algorithm minimizes not only...

In a distributed real-time system, uneven task arrivals temporarily overload some nodes while leaving others idle or underloaded. Consequently, some tasks may miss their deadlines even if the overall system has the capacity to meet the deadlines of all tasks. An effective load sharing (LS) scheme is proposed as a solution to this problem. Upon arrival of a task at a node, the node determines whether or not the node can complete the task in time under the minimum--laxity--first--served policy. If the task cannot be guaranteed or if guarantees of some other tasks are to be violated due to the addition of this task to the existing schedule, the node looks up the list of loss--minimizing decisions , and determines the best node among a set of nodes in its physical proximity, called its buddy set , to which the task(s) may be transferred. This list of decisions is periodically updated using Bayesian decision analysis and prior/posterior state distributions. These probability distributions a...

by Karen Marie Tracey A processor sharing system allows busy users in a networked environment to take advantage of the processing power of idle machines. Experimental systems have demonstrated the usefulness of the concept, but processor sharing has yet to achieve widespread acceptance. Simplistic sharing policies and programming difficulty are two factors that combine to limit the use and acceptance of processor sharing. Group management, a new approach to processor sharing presented here, addresses these problems. Group management is designed to support processor sharing applications that consist of multiple cooperating tasks. This new approach provides a framework for the development of sophisticated sharing policies that can support a variety of applications with differing execution characteristics. In addition, it provides services to programmers that ease the task of writing programs for, and running programs in, a processor sharing environment. This dissertation introduces group...

Networking computers has provided a great potential for performance improvement of job execution through dynamic load balancing. In effect, the literature on this subject concludes that dynamic load balancing algorithms can significantly reduce jobs waiting time. The application of these encouraging results onto real systems has proven to be difficult due to the complexity, lack of support, and lack of adequate software tools for constructing distributed load balancing systems This paper presents the design, implementation and a performance analysis of a dynamic load balancing system. The performance evaluation is undertaken using an Office Information System (OIS) as a case study. The dynamic load balancing prototype has been developed using a distributed programming model which supports dynamic configuration. It is argued that this latter exhibits capabilities that are essential for a system operating in a changing and evolving environment. The overall system has been implemented an...

In a distributed-computing environment, it is important to ensure that the processor workloads are adequately balanced. Among numerous load-balancing algorithms, a unique approach due to Das and Prasad defines a symmetric broadcast network (SBN) that provides a robust communication pattern among the processors in a topology-independent manner. In this paper, we propose and analyze three novel SBN-based load-balancing algorithms, and implement them on an SP2. A thorough experimental study with Poisson-distributed synthetic loads demonstrates that these algorithms are very effective in balancing system load while minimizing processor idle time. They also compare favorably with several other existing load-balancing techniques. Additional experiments performed with real data demonstrate that the SBN approach is effective in adaptive computational science and engineering applications where dynamic load balancing is extremely crucial. Key words: Hypercube, job migration, load balancing, mes...

In this paper we survey some techniques that are useful for proving the optimality of control procedures for networks. These techniques are based on the comparison of sample path realizations under different procedures in order to show that one performs better than the other. Although sample path analysis is capable of solving only a subset of problems that arise in the design of optimal control procedures, it typically requires a minimal set of assumptions regarding the workload and provides considerable insight into the behavior of different procedures. The other main technique based on Markov decision processes usually requires Markovian assumptions and provides less insight. We review different ways of comparing sample paths based on the ideas of vector majorization and set dominance. Applications of these two types of comparisons are drawn from problems in routing, internetwork flow control, and scheduling of real-time data. These applications attempt to address the problems of h...

Network Analysis Without Exponentiality Assumptions by Mor Harchol-Balter Doctor of Philosophy in Computer Science University of California at Berkeley Professor Manuel Blum, Chair Theoretical analysis of networks often relies on exponentiality assumptions which are unrealistic but necessary for mathematical tractability. In some cases these assumptions are inconsequential, but in others their adoption may lead to invalid results.