Search

We propose new burst assembly schemes and fast reservation (FR) protocols for Optical Burst Switched (OBS) networks that are based on traffic prediction. The burst assembly schemes aim at minimizing (for a given burst size) the average delay of the packets incurred during the burst assembly process, while the fast reservation protocols aim at further reducing the end-to-end delay of the data bursts. The burst assembly techniques use a linear prediction filter to estimate the number of packet arrivals at the ingress node in the following interval, and launch a new burst into the network when a certain criterion, different for each proposed scheme, is met. The fast reservation protocols use prediction filters to estimate the expected length of the burst and the time needed for the burst assembly process to complete. A Burst Header Packet (BHP) packet carrying these estimates is sent before the burst is completed, in order to reserve bandwidth at intermediate nodes for the time interval the burst is expected to pass from these nodes. Reducing the packet aggregation delay and the time required to perform the reservations, reduces the total time needed for a packet to be transported over an OBS network and is especially important for real-time applications. We evaluate the performance of the proposed burst assembly schemes and show that a number of them outperform the previously proposed timer-based, length-based and average delay-based burst assembly schemes. We also look at the performance of the fast reservation (FR) protocols in terms of the probability of successfully establishing the reservations required to transport the burst.

We consider information aggregation as a method for reducing the information exchanged in a Grid network and used by the resource manager in order to make scheduling decisions. In this way, information is summarized across nodes and sensitive or detailed information can be kept private, while resources are still publicly available for use. We present a general framework for information aggregation, trying to identify issues that relate to aggregation in Grids. In this context, we describe a number of techniques, including single point and intra-domain aggregation, define appropriate grid-specific domination relations and operators for aggregating static and dynamic resource information, and discuss resource selection optimization functions. The quality of an aggregation scheme is measured both by its effects on the efficiency of the scheduler's decisions and also by the reduction it brings on the amount of resource information recorded, a tradeoff that we examine in detail. Simulation experiments demonstrate that the proposed schemes achieve significant information reduction, either in the amount of information exchanged, or in the frequency of the updates, while at the same time maintaining most of the value of the original information as expressed by a stretch factor metric we introduce.

We examine a task scheduling and data migration problem for grid networks, which we refer to as the Data Consolidation (DC) problem. DC arises when a task concurrently requests multiple pieces of data, possibly scattered throughout the grid network, that have to be present at a selected site before the task's execution starts. In such a case, the scheduler and the data manager must select (i) the data replicas to be used, (ii) the site where these data will be gathered for the task to be executed, and (iii) the routing paths to be followed; this is assuming that the selected datasets are transferred concurrently to the execution site. The algorithms or policies for selecting the data replicas, the data consolidating site and the corresponding paths comprise a Data Consolidation scheme. We propose and experimentally evaluate several DC schemes of polynomial number of operations that attempt to estimate the cost of the concurrent data transfers, to avoid congestion that may appear due to these transfers and to provide fault tolerance. Our simulation results strengthen our belief that DC is an important problem that needs to be addressed in the design of data grids, and can lead, if performed efficiently, to significant benefits in terms of task delay, network load and other performance parameters.

A key problem in Grid networks is how to efficiently manage the available infrastructure, in order to satisfy user requirements and maximize resource utilization. This is in large part influenced by the algorithms responsible for the routing of data and the scheduling of tasks. In this paper, we present several multi-cost algorithms for the joint scheduling of the communication and computation resources that will be used by a Grid task. We propose a multi-cost scheme of polynomial complexity that performs immediate reservations and selects the computation resource to execute the task and determines the path to route the input data. Furthermore, we introduce multi-cost algorithms that perform advance reservations and thus also find the starting times for the data transmission and the task execution. We initially present an optimal scheme of non-polynomial complexity and by appropriately pruning the set of candidate paths we also give a heuristic algorithm of polynomial complexity. Our performance results indicate that in a Grid network in which tasks are either CPU- or data-intensive (or both), it is beneficial for the scheduling algorithm to jointly consider the computational and communication problems. A comparison between immediate and advance reservation schemes shows the trade-offs with respect to task blocking probability, end-to-end delay and the complexity of the algorithms.

We present and theoretically and experimentally analyze a Quality of Service (QoS) framework for Grids that provides (i) deterministic delay bounds to Guaranteed Service (GS) users and (ii) fair sharing of resources to Best Effort (BE) users. The framework adopts concepts from Data Networks and applies them in the Grid environment. We initially describe the proposed framework assuming that task computational workloads are known (or can be estimated), and then provide extensions for the more realistic case where we have no a-priori knowledge of the task workloads. Task migration across multiple resources is also examined in this context. We also look at a specific implementation of the proposed QoS scheme, where we distinguish computational resources, based on the type of users (GS or BE) they serve and the priority they give to each type. We validate experimentally the proposed QoS framework for Grids, verifying that it satisfies the delay guarantees promised to GS users and provides fairness among BE users, while simultaneously improving performance in terms of deadlines missed and resource utilization. In our simulations, data from a real Grid Network are used.