On the Optimization of Load Balancing in Distributed Networks in the Presence of Delay

Abstract

Summary.In large-scale distributed-computing environments, there are a number of inherent time-delay factors that can seriously alter the expected performance of load-balancing (or scheduling) policies that do not account for such delays. This situation arises, for example, in systems for which each computational element (CE) is connected by means of a shared broadband communication medium. The delays, in addition to being large, fluctuate randomly as a result of uncertainties in the network condition and uncertainty in the size of the loads to be transferred between the CEs. The performance of such distributed systems under any load-balancing policy is therefore stochastic in nature and must be assessed in a probabilistic sense. Moreover, the design of load-balancing policies that best suit such delay-infested environments must also be carried out in a statistical framework. Indeed, it has been observed in recent simulation-based and experimental studies that the delay, especially that corresponding to transferring large loads between CEs, can cause the system to fall into a mode whereby CEs unnecessarily exchange loads back and forth. This results in an undesirable situation where loads continue to be in transit. It has been observed that the notion of limiting the number of balancing instants in scheduling algorithms can be used to customize certain load-balancing policies to random-delay environments. But the effectiveness of this strategy depends on the decision as to when and how often such load-balancing cycles are to be executed. If a system starts from a certain unbalanced state, we would expect that there would be an optimal time at which balancing must be executed so as to minimize the overall completion time of a workload. In particular, the load-balancing instant should be large enough to ensure that all CEs have updated knowledge of the state of the system. Yet, the balancing instant should be soon enough so that the system does not remain in an unbalanced state for an excessively prolonged time whereby allowing some of the CEs to be idle. It has also been noticed that an additional strategy for mitigating the effects of delay in load balancing is to reduce the gain, or strength, of the load-balancing policy. That is, for each overloaded CE, only a fraction of the load to be transferred is actually transferred to other CEs. Thus, in view of these unique phenomena that occur in load balancing in delay-infested networks, optimizing the performance over the inter-balancing times and the load-2 Authors Suppressed Due to Excessive Length balancing gains becomes an important problem. In this Chapter, the performance of a single-instant load-balancing strategy on a distributed physical system is studied both theoretically and experimentally. The experiments are performed on an inhouse wireless distributed testbed as well as a Monte-Carlo simulation tool. The theoretical analysis is carried out based on the concept of regeneration in stochastic processes. In particular, a probabilistic analysis of the queu...

Load balancing in the presence of random node failure and recovery

“…This is a more plausible way to calculate the excess load of a node in a heterogeneous computing environment as compared to our earlier method that did not consider the processing speeds of the nodes [8][9][10][11]. Clearly, with the inclusion of the processing speeds of the nodes, a slower (faster) node has a larger (smaller) excess load than what it would have had under our earlier definition.…”

“…It has been shown in our previous works that in a delay infested system, the division of the load based on the nodes' processing speeds alone does not yield a minimal average overall completion time [8,10]. Specifically, at time t = 0, every node computes its excess load by comparing its load to the average over all load of the system.…”

Distributed Load Balancing in the Presence of Node Failure and Network Delays

“…To test the performance of the newly proposed load-balancing policy denoted as lb2, a comparative experiment was performed between this policy and the strategy adopted by the stochastic model introduced in Dhakal et al, 2004) and referred to as lb1 (1). The experiments were conducted over PlanetLab(http://www.planet-lab.org) and the initial settings and parameters are shown in Table 1.…”

“…Similarly, in Ghanem et al, 2004b), Monte-Carlo studies and real-time experiments were conducted over the wireless LAN to study one-shot dynamic load-balancing scheme for distributed systems. Moreover, a regenerationtheory-based mathematical model, represented by a set of coupled integro-difference equations, for one-shot load-balancing policy as applied to a twonodes distributed system was presented (Dhakal, 2003;Ghanem et al, 2004b;Dhakal et al, 2004).…”

“…The occurrence of load-balancing instances, i.e. when to trigger the load distribution strategy, is not discussed in this paper since it is possible to employ one-shot load-balancing, multiple balancing instances, or any other scheme that would be suitable for the system at hand (Ghanem, 2004;Ghanem et al, 2004b;Dhakal et al, 2005;Dhakal et al, 2004;Dhakal, 2003). Furthermore, exchanged information about the state of the nodes is still assumed to occur frequently with additional information to be discussed subsequently.…”