Bernard Toursel

A Java CPU calibration tool for load balancing in distributed applications

This article presents a method for evaluating the CPU power, independently from the system used, in heterogeneous networks of workstations. It is based on the use of Java language in order to ensure application portability and more particularly on the mechanism of thread CPU processing time measurement introduced in the version 1.5 of Sun Java. That tool will be integrated into the load balancing mechanism which is totally written in Java and that we developed in the LIFL project ADAJ. We show how to evaluate the potential power of the CPU with a software totally written in Java. Moreover, we will justify the results provided by our approach. We will also analyse the exploitation of the calibration tool in order to improve the execution time of parallel and distributed applications in the context of load balancing in a network of workstations.

Embedding trees in massively parallel computers

Abstract Trees are a useful class of computational structures. In parallel processing, mapping problems arise when the tree structure differs from the processor interconnection of the parallel computer. The MasPars MP-2 is a massively parallel computer where processors are interconnected via the X-Net neighborhood two-dimensional mesh and the global multistage crossbar router network. We present novel mapping schemes for trees on the MasPars MP-2 two-dimensional mesh and on the two-dimensional mesh together with the MasPars MP-2 multistage crossbar network. Appropriate algorithms of the mapping scheme on the two-dimensional mesh (or grid) are presented and are shown to be superior over known mappings on square arrays (or grids).

A Framework for Desktop GRID Applications: CCADAJ

We propose a new component framework over a middleware platform for grid computing: DG-ADAJ (desktop grid-adaptive distributed application in Java). Our platform allows having a single system image (SSI) of the grid platform. It gives a special mechanism at middleware level which assures dynamic and automatic adaptation to variations of computation methods and execution platform. DG-ADAJ gives also specific mechanisms based on control components which helps a users to build component-based parallel/distributed applications. We use the CCA (common component architecture) as a component architecture model for our framework. CCADAJ (CCA-ADAJ) is a layer above the DG-ADAJ environment, which helps user to build his parallel/distributed applications by assembling components in a transparent way and which exploits the parallelism of the environment

Load balancing in the SOAJA web service platform

The aim of the service oriented adaptative Java applications (SOAJA) project is to develop a service-oriented infrastructure, which enables efficient running of Java applications in complex, networked computing Grid environments. The SOAJA environment provides components and services for static and dynamic load balancing based on Java object observation. SOAJA can be used to design large scale computing tasks to be executed based on idle time of processor nodes. Java distributed applications consist of parallel objects which SOAJA allocates to Grid nodes at runtime. In this paper, we present mechanisms and algorithms for automatic placement and adaptation of application objects, in response to evolution of resource availability. These mechanisms enable to control the granularity data processing and distribution of the application on the Grid platform.

Embedding tree structures in massively parallel computers

Trees are an important class of computational structures due to their favoumble property of logarithmic path from the root to any leaf element. Since simple and regular interconnections among nodes are needed, they are easily manipulated. Mapping problems arise when the tree structure differs from the processor interconnection of the pamllel computer. The MasPar’s MP-2 is a massively parallel computer where processors are interconnected via the X-Net neighborhood two-dimensional mesh and the globul multistage crossbar router network. We present novel mapping schemes for trees on the MasPar’s MP-2 twodimensional mesh and on the two-dimensional mesh together with the MasPar’s MP-2 multistage crossbar network. Appropriate algorithms of the mapping scheme on the two-dimensional mesh (or grid) are presented and are shown to be superior over known mappings on square arrays (or grids). keywords: massively parallel systems, interconnection networks, mapping scheme, tree structures, distributed configuration algorithm, evaluation.

Java application execution in heterogeneous and multi-user systems

This article presents an inter-application load balancing mechanism meant to improve the efficiency of execution for Java parallel and distributed applications in a network of workstations. One difficulty of such architectures is dealing with heterogeneity: workstations have different powers and operating systems. Moreover, the workstations are generally executing several processes, corresponding to an external load, outside the JVM, as in the case of multi-user systems. Consequently, heterogeneity and resource sharing are features included in the design of the dynamic load balancing in ADAJ. Two new Java tools, of calibration and of load measure, are proposed and used to extend the previously load balancing approach applicable on a dedicated and homogeneous environment.

Performance Of Very High Frequencies Networks

Very high frequencies technology is already well established for LAN’ and W A P . This paper deals with the use of very high frequencies for message passing communications in massively parallel computers. We present some experimental results which validate the possible use of simultaneous communications between computers through a single waveguide. Furthermore, an original proposition of waveguides interconnections which is well suited to a large interconnection network and providing both a high connectivity degree and a very shod diameter is discussed. The performance analysis using analytical queuing models of the network is shown; we present the network bandwidth and latency as a function of system load and architectural parameters of the machine. keywords: interconnection networks, very high frequencies, bandwidth, performance analysis.

A Time-Bounded Binding Access with Low Task-Switching Cost in an OR-Parallel Prolog

A new binding scheme for OR-multisequential execution of prolog programs is presented. The proposed scheme is a shared binding environment one, between the binding arrays scheme which has a constant time dereferencing and the hash-windows scheme which has low task-switching. The proposed scheme enables to have a bounded dereferencing time and a constant task-switching time. We show then this scheme allows to use various scheduling strategies and in particular leftmost strategy which enables to handle efficiently speculative work.

An Efficient Binding Management in OR-Parallel Model

A new binding scheme is presented which allows a depth first scheduling strategy in order to manage efficiently side effect predicates. The dereferencing mechanism is based on a set of N key registers associated to each processor. Each key register Ki selects conditional bindings made by processor Pi and which are valid for a given processor Pk.