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...

Abstract not found

We address the problem of performance prediction for parallel programs executed on clusters of heterogeneous workstations on which resource contention is present. We develop a methodology for the construction of performance models whose analysis allows the estimation of the execution time of these programs. We use Timed Petri Nets to represent the behavior of parallel programs, and a contention model based on queueing theory to quantify the effects of resource contention on the execution time of the application processes. Our methodology is demonstrated through the construction of the model of an example program, which is also used to validate the predictions against measured execution times obtained by executing the program on two different clusters of workstations. 1 Introduction Clusters of workstations are increasingly being used as costeffective parallel computing platforms, but their effective exploitation requires consideration of many of the research problems already addresse...

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...

Load sharing is a technique to improve the performance of distributed systems by distributing the system workload from heavily loaded nodes to lightly loaded nodes in the system. Previous studies have considered two adaptive load sharing policies: sender-initiated and receiver-initiated. In the sender-initiated policy, a heavily loaded node attempts to transfer work to a lightly loaded node and in the receiver-initiated policy a lightly loaded node attempts to get work from a heavily loaded node. Most of the previous studies have assumed the first-come/first-served (FCFS) node scheduling policy; furthermore, analyses and simulations in these studies have been done under the assumption that the job service times are exponentially distributed and the job arrivals form a Poisson process. The goal of this paper is to fill the void in the existing literature. We study the impact of these assumptions on the performance of the sender-initiated and receiverinitiated policies. We consider thre...

Distributed-controlled dynamic load balancing algorithms are known to have several advantages over centralised algorithms such as scalability, and fault tolerance. Distributed implies that the control is decentralised and that a copy of the algorithm (called a scheduler) is replicated on each host of the network. However, distributed control also contributes to the lack of global goals and lack of coherence. This paper presents a new algorithm called DGP (Decentralised Global Plans) that addresses the problem of coherence and coordination in distributed dynamic load balancing algorithms. The DGP algorithm is based on a strategy called Global Plans (GP), and aims at maintaining all computational loads of a distributed system within a band called D. The rationale for the design of DGP is to allow each scheduler to consider the actions of its peer schedulers. With this level of coordination, the schedulers can act more as a coherent team. This new approach first explicitly specifies a glo...

Distributed computing systems consist of computers interconnected by communications links. In such systems, statistical fluctuations in job arrival and service patterns cause episodes of load imbalance during which some computers are lightly loaded while others are simultaneously overloaded. Load sharing is the process of transferring jobs from overloaded to underloaded computers to improve overall system performance. Load sharing is implemented by means of load sharing algorithms. However, since little has been known about the temporal characteristics of load imbalance, various pitfalls arise in the operation of load sharing algorithms which prevent full realization of the potential benefits of load sharing. In this paper, we present a stochastic analysis of time durations of load imbalance. The notions of Transfer Pair , and Load Sharing Window are introduced. Using first passage times, a general expression for the probability distribution function of the Load Sharing Wi...

Multithreading is a promising approach to address the problems inherent in multiprocessor systems, such as network and synchronization latencies. Moreover, the benefits of multithreading are not limited to loop-based algorithms but apply also to irregular parallelism. EARTH - Efficient Architecture for Running THreads, is a multithreaded model supporting fine-grain, non-preemptive threads. This model is supported by a C-based runtime system which provides the multithreaded environment for the execution of concurrent programs. This thesis describes the design and implementation of a set of dynamic load balancing algorithms, and an in-depth study of their behavior with divide-and-conquer, regular, and irregular classes of applications. The results described in this thesis are based on EARTH-SP2, an implementation of the EARTH program execution model on the IBM SP-2, a distributed memory multiprocessor system. The main results of this study are as follows: ffl A randomizing load balance...

This paper presents the performance evaluation of a coherent distributed dynamic load balancing algorithm called the DGP (Decentralised Global Plans) algorithm. The performance evaluation is underpinned by a research methodology integrating a simulation-based performance evaluation with a prototype-based performance evaluation. The DGP algorithm addresses the issue of coherence and coordination in distributed dynamic load balancing algorithms. The DGP algorithm is based on a strategy called Global Plans (GP). This new approach first explicitly specifies a global goal and then design a strategy around this global goal such that each scheduler (1) takes into account local decisions made by other schedulers; (2) takes into account the effect of its local decisions on the overall system and (3) ensures load balancing. A simulation-based performance evaluation of DGP with two other well-known dynamic load balancing algorithms published in the literature shows that DGP performs consistently ...

This paper describes a dynamic, symmetrically-initiated load sharing scheme which adapts to changing load condition by varying the algorithm's dependency on system's status information. The scheme is hybrid in that it relies on a a fully distributed algorithm when the system is heavily loaded, but resorts to a centrally coordinated location policy when parts of the system become idle. The simplicity of the algorithms proposed makes it possible to use a centralized component without incurring in scalability problems and presenting instabilities. Both algorithms are very lightweight and do not need any tuning of parameters, so that they are extremely easy to implement to the point that an inexpensive hardware implementation of the centralized component is capable of handling millions of requests per second. Simulations show that the hybrid approach outperforms existing dynamic algorithms under all load conditions and task generation patterns, it is weakly sensitive to processing overhead...