Spiridon Pulidas, Don Towsley, and John A. Stankovic. Imbedding gradient estimators in load balancing algorithms. In 8th International Conference on Distributed Computing Systems, pages 482--490, June 1988.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....and RHODOS [27] Preemptive migration has been found to be very effective for load balancing in MOSIX, 7] 8.2.2 Studies Migration policies have been studied much more in simulations and analyses. Most of these studies consider only non preemptive policies, for example, 53] 87] 15] 96] [61], 42] 9] 25] 50] 55] 94] 97] 29] 22] Only a few studies have considered preemptive migration policies, in addition to non preemptive ones: specifically, 23] 41] 47] and [12] However these studies disagree on the effectiveness of preemptive migration. Eager, Lazowska, and ....

....migration is considered. Even so, the state space of the CTMC model is still very large, so further simplifying assumptions are made to reduce the state space. 9.1. 2 Exponential assumptions in simulation papers Exponential assumptions are also made in strictly simulation papers, e.g. 42] [61], 87] 25] 53] 94] 17] and many others. Most of these papers do not justify their use of exponentiality assumptions. One of these, 42] acknowledge that an exponential workload many not be typical, but state that they believe and exponential workload is accurate enough for the purpose of ....

[Article contains additional citation context not shown here]

Spiridon Pulidas, Don Towsley, and John A. Stankovic. Imbedding gradient estimators in load balancing algorithms. In 8th International Conference on Distributed Computing Systems, pages 482--490, June 1988.

....aspects of the development of the system: the dynamic load balancing algorithm used; and the distributed programming model used to implement the algorithm. 3. 1 The Load Balancing Algorithm There is an extensive literature on load balancing in loosely coupled distributed systems (see for example [FYN88, LK87, PTS88]) Typically, a load balancing algorithm has three types of processes [ZF87] 1. A load information process. This determines the type of information that makes a load index, and the way such information is communicated to other schedulers (for example, broadcasting, focussed addressing, ....

S Pulidas, D Towsley, and J A Stankovic. Imbedding gradient estimators in load balancing algorithms. Proceedings of the Eighth International Conference on Distributed Computing Systems, pages 482 -- 490, June 1988.

....param etres de l algorithme sont mis a jour dynamiquement en fonction de l etat courant du syst eme [CK88] La valeur du seuil T , des politiques de transfert bas ees sur les seuils, peut par exemple etre d etermin ee dynamiquement. Une d etermination adaptative du seuil a et e propos ee dans [PTS88] La technique propos ee est d estimer p eriodiquement la charge moyenne du syst eme et de mettre a jour le seuil si la diff erence entre la charge moyenne du syst eme et la charge locale d epasse une certaine valeur. Une m ethode des ench eres adaptative consistant a ajuster dynamiquement la ....

S. Pulidas, D. Towsley, and J. A. Stankovic. Imbedding gradient estimators in load balancing Algorithms. In Proc. of the 8th Int. Conf. on Distributed Computing Systems, San Jose, California, pages 482--490, Jun 1988.

....so the system is balanced. The output set is (4, 5, 4, 5, 3) and the global table of allocations can be represented by (2, 2, 1) 1, 4, 3) 3.1. The Decentralised Global Plans Algorithm There is an extensive literature on load balancing in loosely coupled distributed systems (see for example [10, 21, 31]) Typically, a load balancing algorithm has three types of policies: 1. A load information policy. This determines the type of information that makes a load index, and the way such information is communicated to other schedulers (for example, broadcasting, focussed addressing, polling) 2. A ....

S Pulidas, D Towsley, and J A Stankovic, "Imbedding Gradient Estimators in Load Balancing Algorithms," Proceedings of the Eighth International Conference on Distributed Computing Systems, pp. 482 - 490, IEEE, June 1988.

....In particular, we need to explain the interpretation of reinforcement learning and the interpretation of load balancing we adopt. Much work has been devoted in the recent years to distributed and adaptive load balancing. One can find related work in the field of distributed computer systems (e.g. Pulidas, Towsley, Stankovic, 1988; Mirchandaney Stankovic, 1986; Billard Pasquale, 1993; Glockner Pasquale, 1993; Mirchandaney, Towsley, Stankovic, 1989; Zhou, 1988; Eager, Lazowska, Zahorjan, 1986) in organization theory and management science (e.g. Malone, 1. In many applications the capacities of the resources are ....

Pulidas, S., Towsley, D., & Stankovic, J. (1988). Imbedding gradient estimators in load balancing algorithms. In Proceedings of the 8th International Conference on Distributed Computer Systems, IEEE, pp. 482--489.

....Experimental systems researchers have focused, for the most part, on the challenging questions of mechanisms; the policies that have been implemented are usually quite simple. However, many researchers have performed analytic studies of various aspects of load sharing balancing policies [1, 8, 10, 25, 27, 28, 29, 31, 38, 39, 41, 46, 51, 55, 58]. This section focuses on research in this vein with results that are applicable in situations involving implementation of processor sharing on a network of general purpose workstations. This includes studies of: the relative performance of various task placement algorithms (Section 2.3.1) and ....

....it is decided that some job should not be run on the local processor, that decision is usually made because the load on the local processor is above some threshold value. The threshold is a tuning parameter; its optimal value is a function of the overall system state, which changes over time. In [46], Pulidas et al. consider the problem of assigning the proper values to threshold parameters. They describe an iterative algorithm where hosts compute local information and 12 exchange it with other hosts in order to appropriately update the threshold value for transferring jobs for remote ....

Pulidas, Spiridon, Towsley, Don, and Stankovic, John A., "Imbedding Gradient Estimators in Load Balancing Algorithms", Proceedings of the 8th International Conference on Distributed Computing Systems, June 1988, pp. 482-490.

....few systems use preemptive migration for load balancing, there have been many simulation studies and analytic models showing the performance benefits of various load balancing strategies. Some of these studies have focused on load balancing by remote execution ( LM82] WM85] CK87] Zho87] PTS88] Kun91] HJ90] ELZ86] others have compared the performance of systems with and without preemptive migration ( ELZ88] KL88] Our work differs from [ELZ88] in both system model and workload description. ELZ88] model a server farm in which incoming jobs have no affinity for a particular ....

Spiridon Pulidas, Don Towsley, and John A. Stankovic. Imbedding gradient estimators in load balancing algorithms. In 8th International Conference on Distributed Computing Systems, pages 482--490, June 1988.

....Although few systems incorporate migration policies, there have been many simulation and analytical studies of various migration policies. Most of these studies have focused on load balancing by remote execution [Livny and Melman 1982] Wang and Morris 1985] Casavant and Kuhl 1987] Zhou 1987] [Pulidas et al. 1988] [Kunz 1991] Bonomi and Kumar 1990] Evans and Butt 1993] Lin and Raghavendra 1993] Mirchandaney et al. 1990] Zhang et al. 1995] Zhou and Ferrari 1987] Ha c and Jin 1990] Eager et al. 1986] Only a few studies address preemptive migration policies [Leland and Ott 1986] Krueger and Livny ....

....age. 2. 2 Why the distribution is critical Many prior studies of process migration assume an exponential distribution of process lifetimes, both in analytical papers [Lin and Raghavendra 1993] Mirchandaney et al. 1990] Eager et al. 1986] Ahmad et al. 1991] and in simulation studies [Kunz 1991] [Pulidas et al. 1988] [Wang and Morris 1985] Evans and Butt 1993] Livny and Melman 1982] Zhang et al. 1995] Chowdhury 1990] The reasons for this assumption include: 1) analytic tractability, and (2) the belief that even if the actual lifetime distribution is in fact not exponential, assuming an exponential ....

Pulidas, S., Towsley, D., and Stankovic, J. A. 1988. Imbedding gradient estimators in load balancing algorithms. In 8th International Conference on Distributed Computing Systems (June 1988), pp. 482--490.

....These techniques are concerned with analyzing a sample path in order to obtain a sample path gradient [17] so as to determine how the control parameter should be changed. Such analyses have been developed and evaluated for network routing [5, 10] and load balancing in distributed systems [6, 13]. Although an interesting area of research, they will not be covered in this paper. The remainder of the paper is organized in the following manner. Section 2 describes two different types of comparisons that are useful in sample path analysis. The next two sections, 3 and 4, present applications ....

S. Pulidas, D. Towsley, J.A. Stankovic. "Imbedding gradient estimators in load balancing algorithms", Proceedings of Eighth International Conference on Distributed Computing, pp. 482-490, June 1988.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC