Len Freeman

Euro-Par 2001 Parallel Processing

A software component framework is one where an application designer programs by composing well understood and tested “components” rather than writing large volumes of not-very-reusable code. The software industry has been using component technology to build desktop applications for about ten years now. More recently this idea has been extended to application in distributed systems with frameworks like the Corba Component Model and Enterprise Java Beans. With the advent of Grid computing, high performance applications may be distributed over a wide area network of compute and data servers. Also “peerto-peer” applications exploit vast amounts of parallelism exploiting the resources of thousands of servers. In this talk we look at the problem of building a component technology for scientific applications. The common component architecture project seeks to build a framework that allows software components runing on a massively parallel computers to be linked together to form wide-area, high performance application services that may be accessed from desktop applications. This problem is far from being solved and the talk will describe progress to date and outline some of the difficult problems that remain to be solved. R. Sakellariou et al. (Eds.): Euro-Par 2001, LNCS 2150, pp. 5–5, 2001. c

Feedback guided dynamic loop scheduling; A theoretical approach

In this paper we review existing loop scheduling algorithms and also describe the feedback-guided dynamic loop scheduling (FGDLS) algorithm that was proposed in Bull et al. (1996) and Bull (1998). The FGDLS algorithm uses a feedback mechanism to schedule a parallel loop within a sequential outer loop. It has been shown to perform well for scheduling problems for which the load associated with the parallel loop changes relatively slowly as the outer sequential loop executes. However the question of convergence of the FGDLS algorithm has remained an open question. In this paper we are able to establish sufficient conditions (essentially requiring that the workload does not change too rapidly with loop iteration count) for the (global) convergence of a continuous analogue of the feedback-guided algorithm.

The design of a performance steering system for component-based grid applications

A major method of constructing applications to run on a computational Grid is to assemble them from components - separately deployable units of computation of well-defined functionality. Performance steering is an adaptive process involving run-time adjustment of factors affecting the performance of an application. This paper presents a design for a system capable of steering, towards a minimum run-time, the performance of a component-based application executing in a distributed fashion on a computational Grid. The proposed performance steering system controls the performance of single applications, and the basic design seeks to separate application-level and component-level concerns. The existence of a middleware resource scheduler external to the performance steering system is assumed, and potential problems are discussed. A possible model of operation is given in terms of application and component execution phases. The need for performance prediction capability, and for repositories of application-specific and component-specific performance information, is discussed. An initial implementation is briefly described.

Performance evaluation of storage formats for sparse matrices in fortran

Many storage formats have been proposed to represent spa- rse matrices. This paper extends to Fortran 95 the performance evaluation of sparse storage formats in Java presented at ICCS 2005, [7]. These experiments have the same set up (almost 200 sparse matrices and matrix-vector multiplication), but now consider the Fortran 95 Sparse BLAS reference implementation.

A static workload balance scheduling algorithm

This article studies a static scheduling method based on workload balancing. An equation is presented for the case when the workload is equally distributed onto all the processors. An efficient load balance scheduling algorithm is developed assuming that the workload has certain properties. Finally, some computational results are given for the product between an upper diagonal matrix and a vector.

AN O∥LOG P) PARALLEL IMPLEMENTATION OF FEEDBACK GUIDED DYNAMIC LOOP SCHEDULING

Feedback Guided Dynamic Loop Scheduling (FGDLS) is a recently proposed dynamic algorithm for loop scheduling. The original algorithm required an O(p) serial computation at each stage to compute the updated loop schedule. In this paper, it is shown that this computation can be implemented in O(log p) operations on p processors

A Theoretical Application of Feedback Guided Dynamic Loop Scheduling

In this paper we briefly describe the Feedback-Guided Dynamic Loop Scheduling (FGDLS) algorithm that was proposed in Bull et al. [2] and Bull [1]. The FGDLS algorithm uses a feedback mechanism, based on measured execution times, to schedule a parallel loop within a sequential outer loop. We compare the FGDLS algorithm with other scheduling algorithms for a simple model problem -- the parallel computation of the inverse of a triangular matrix.

Feedback Guided Dynamic Integral Partition

In this article we introduce a new iterative method for integral partition called the feedback guided dynamic integral partition (FGDIP) algorithm. The problem to study is the partition of a definite integral into p identical sub-integrals. The method generates iteratively a sequence of integral bounds by re-balancing the previous integral partition to achieve a better one. A simple convergence condition is also proposed. Experimental results show that the proposed method FGDIP achieves better performance than the classical Newtons method

Static workload balance scheduling; continuous case

This article studies a static scheduling method based on workload balancing in the continuous case. An equation is presented for the case when the workload, as continuous-function is equally distributed onto processors based on integrals. A sufficient condition is also established for the fully covering property. Finally, some computational results are given to prove that the continuous case is better that the discrete case.

Integrating Dynamic Memory Placement with Adaptive Load-Balancing for Parallel Codes on NUMA Multiprocessors

This Paper describes and evaluates a system of dynamic memory migraton for codes executing in a Non-Uniform Memory Access environment. This system of migration applies information about the load-imbalance within a workload in order to determine the affinity between threads of the application and regions of memory. This information then serves as the basis of migration decisions, with the object of minimising the NUMA distance between code and the memory it accesses. Results are presented which demonstrate the effectiveness of this technique in reducing the runtime of a set of representative HPC kernels.