Stavros Souravlas

A pipeline technique for dynamic data transfer on a multiprocessor grid

This paper describes a pipeline technique which is used to redistribute data on a multiprocessor grid during runtime. The main purposes of the algorithm are to minimize the data transfer time, prevent congestion on the ports of the receiving processors, and minimize the number of idle processors. One of the key ideas for this algorithm is the creation of processor classes, firstly introduced by Desprez et al. [IEEE Transactions on Parallel and Distributed Systems 9(2):102 (1998).] Based on the idea of classes, we create the pipeline tasks used to organize the redistribution of data. Our experimental results show that this pipeline technique can significantly reduce the amount of time required to complete a dynamic data transfer task.

Binary-Tree Based Estimation of File Requests for Efficient Data Replication

Recently, data replication has received considerable attention in the field of grid computing. The main goal of data replication algorithms is to optimize data access performance by replicating the most popular files. When a file does not exist in the node where it was requested, it necessarily has to be transferred from another node, causing delays in the completion the file requests. The general idea behind data replication is to keep track of the most popular files requested in the grid and create copies of them in selected nodes. In this way, more file requests can be completed over a period of time and average job execution time is reduced. In this paper, we introduce an algorithm that estimates the potential of the files located in each node of the grid, using a binary tree structure. Also, the file scope and the file type are taken into account. By potential of a file, we mean its increasing or decreasing demand over a period of time. The file scope generally refers to the extent of the group of users which are interested or potentially interested in a file. The file types are divided into read and write intensive. Our scheme mainly promotes the high-potential files for replication, based on the temporal locality principle. The simulation results indicate that the proposed scheme can offer better data access performance in terms of the hit ratio and the average job execution time, compared to other state-of-the-art strategies.

Trends in data replication strategies: a survey

ABSTRACT Data Grids allow many organisations and individuals to share their data across long-distance areas. Nowadays, a huge amount of data is produced in all scientific fields and to enhance collaboration and data sharing, it is necessary to make this data available to as many nodes of the grid as possible. Data replication is the technique used to provide this availability. Moreover, it improves access time and reduces the bandwidth used. Recently, data replication has received considerable attention and several new algorithms have been developed. This article provides an overview of the state-of-the-art techniques of data replication. We identify the advantages and disadvantages of these strategies and discuss about their performance. Graphical Abstract

A message passing strategy for array redistributions in a torus network

The array redistribution problem occurs in many important applications in parallel computing. In this paper, we consider this problem in a torus network. Tori are preferred to other multidimensional networks (like hypercubes) due to their better scalability (IEE Trans. Parallel Distrib. Syst. 50(10), 1201–1218, [2001]). We present a message combining approach that splits any array redistribution problem in a series of broadcasts where all sources send messages of the same size, thus a balanced traffic load is achieved. Unlike existing array redistribution algorithms, the scheme introduced in this work eliminates the need for data reorganization in the memory of the source and target processors. Moreover, the processing of the scheduled broadcasts is pipelined, thus the total cost of redistribution is reduced.

On further reducing the cost of parallel pipelined message broadcasts

In this paper we extend our previous work that focused on reducing the transmission cost of parallel pipelined messages distributed in a block-cyclic fashion. We apply the same transmission strategy but we aim to reduce index computation overheads. More specifically, we show how to reduce the computations required to define the interprocessor communication cost and we introduce a more efficient use of memory, based on indices.

Verification and Performance Evaluation of Parallel Pipelined Communications Using Petri Nets

The growing use of multiprocessing systems has given rise to the necessity of modeling, verifying, and evaluating their performance, in order to fully exploit hardware [9], [15], [16]. The Petri Net (PN) formalism is a suitable tool for modeling parallel systems, because of basic characteristics of these systems, like parallelism and process synchronization. The system under study can be evaluated by means of generating and analyzing a set of processes. In addition [11], the PN formalism allows the incorporation of more details of the real system into the model. Examples of such details include the study of contentions for shared resources (like memory) and the study of blocked processes. In this paper, PN are considered as a modeling framework to verify and study the performance of parallel pipelined communications. The main strength of the pipelines is that, if organized in a proper way, they lead to overlapping of computation, communication, and read/write costs that incur in parallel processing ([7], [1], [14]). The PN model presented in this paper, accurately captures the behavior of a pipeline based parallel communication system. The model considers parallelization, message scheduling, and message classification, while it is proven to be free of deadlocks and contentions. Also, the model is characterized by symmetry, and thus it can be extended for large and complex systems.

On minimizing memory and computation overheads for binary-tree based data replication

Data replication is used to track the most popular files (i.e., the ones with most requests) and replicate them in selected nodes. In this way, more requests for such popular files can be completed over a period of time and bandwidth consumption is reduced, since these files do not need to be transferred from remote nodes. In this article, we extend our previous work [1] to make it more efficient in terms of memory and total computation cost, so that it becomes more efficient and suitable for larger grids. To reduce the memory costs, we present a centralized strategy which estimates the potential for selected batches of files. The computations required for these estimations are executed in a pipelined way, so their cost is also reduced.

Network Load Balancing Using Modular Arithmetic Computations

Load-balanced routing has attracted considerable attention, especially in the recent years, where huge data volumes are carried over the computer networks. It is particularly important for non-all-to-all networks, where there is no direct communication between all the nodes of the network.Telecommunication and network systems constitute complex dynamic systems with an ever-increasing number of users and network services. It has become apparent that, new routing demands can not be easily satisfied by conventional routing methods. Thus, intelligent optimization methods (e.g., nature-inspired methodologies) have arisen to improve network efficiency.This paper presents a computational method that is based on modular arithmetic for achieving dynamic load balancing on data networks. The proposed algorithm organizes the overall communication into equal-sized packets, it divides the communication into a series of communication steps between the network nodes, and performs packet transfer. The last section includes discussion on the main costs each network routing operation inures: the data movement cost, the load information cost and the data reordering cost.

On the Detection of Overlapped Network Communities via Weight Redistributions

A community is an important attribute of networking, since people who join networks tend to join communities. Community detection is used to identify and understand the structure and organization of real-world networks, thus, it has become a problem of considerable interest. The study of communities is highly related to network partitioning, which is defined as the division of a network into a set of groups of approximately equal sizes with minimum number of edges. Since this is an NP-hard problem, unconventional computation methods have been widely applied. This work addresses the problem of detecting overlapped communities (communities with common nodes) in weighted networks with irregular topologies. These communities are particularly interesting, firstly because they are more realistic, i.e., researchers may belong to more than one research community, and secondly, because they reveal hierarchies of communities: i.e., a medical community is subdivided into groups of certain specialties. Our strategy is based on weight redistribution: each node is examined against all communities and weights are redistributed between the edges. At the end of this process, these weights are compared to the total connectivity of each community, to determine if overlapping exists.

Scheduling array redistribution with virtual channel support

This paper addresses the well-known problem of redistributing data arrays over a multiprocessor network. The block-cyclic redistribution techniques found in the literature deal effectively with this problem, based on the assumption that there is a direct link between all the processors of the network. Most of these techniques aim at reducing the number of messages and the total redistribution cost. However, an application of these techniques on non-all-to-all communication networks, like tori, shows that these techniques suffer long delays. In this work, we try to solve the general block-cyclic redistribution problem on non-all-to-all networks, by grouping the messages into well-defined classes, and transferring them with the support of a well-specified number of virtual channels.