MPI (Message Passing Interface) is a specification for a standard library for message passing that was defined by the MPI Forum, a broadly based group of parallel computer vendors, library writers, and applications specialists. Multiple implementations of MPI have been developed. In this paper, we describe MPICH, unique among existing implementations in its design goal of combining portability with high performance. We document its portability and performance and describe the architecture by which these features are simultaneously achieved. We also discuss the set of tools that accompany the free distribution of MPICH, which constitute the beginnings of a portable parallel programming environment. A project of this scope inevitably imparts lessons about parallel computing, the specification being followed, the current hardware and software environment for parallel computing, and project management; we describe those we have learned. Finally, we discuss future developments for MPICH, including those necessary to accommodate extensions to the MPI Standard now being contemplated by the MPI Forum. 1

We describe a methodology for developing high performance programs running on clusters of SMP nodes. Our methodology is based on a small kernel (SIMPLE) of collective communication primitives that make e cient use of the hybrid shared and message passing environment. We illustrate the power of our methodology by presenting experimental results for sorting integers, two-dimensional fast Fourier transforms (FFT), and constraint-satis ed searching. Our testbed is a cluster of DEC AlphaServer 2100 4/275 nodes interconnected by anATM switch.

This paper describes current activities of the MPI-2 Forum. The MPI-2 Forum is a group of parallel computer vendors, library writers, and application specialists working together to define a set of extensions to MPI (Message Passing Interface). MPI was defined by the same process and now has many implementations, both vendor-proprietary and publicly available, for a wide variety of parallel computing environments. In this paper we present the salient aspects of the evolving MPI-2 document as it now stands. We discuss proposed extensions and enhancements to MPI in the areas of dynamic process management, one-sided operations, collective operations, new language binding, real-time computing, external interfaces, and miscellaneous topics. 1 Introduction During 1993 and 1994, a group of parallel computer vendors, library writers, and application scientists met regularly to define a standard interface for message-passing libraries. The result of this effort was MPI (Message-Passing Interfa...

Routing with time windows (VRPTW) has been an area of research that have attracted many researchers within the last 10 { 15 years. In this period a number of papers and technical reports have been published on the exact solution of the VRPTW. The VRPTW is a generalization of the well-known capacitated routing problem (VRP or CVRP). In the VRP a eet of vehicles must visit (service) a number of customers. All vehicles start and end at the depot. For each pair of customers or customer and depot there is a cost. The cost denotes how much is costs a vehicle to drive from one customer to another. Every customer must be visited exactly ones. Additionally each customer demands a certain quantity of goods delivered (know as the customer demand). For the vehicles we have an upper limit on the amount of goods that can be carried (known as the capacity). In the most basic case all vehicles are of the same type and hence have the same capacity. The problem is now for a given scenario to plan routes for the vehicles in accordance with the mentioned constraints such that the cost accumulated on the routes, the xed costs (how much does it cost to maintain a vehicle) or a combination hereof is minimized. In the more general VRPTW each customer has a time window, and between all pairs of customers or a customer and the depot we have a travel time. The vehicles now have to comply with the additional constraint that servicing of the customers can only be started within the time windows of the customers. It is legal to arrive before a time window \opens" but the vehicle must wait and service will not start until the time window of the customer actually opens. For solving the problem exactly 4 general types of solution methods have evolved in the literature: dynamic programming, Dantzig-Wolfe (column generation), Lagrange decomposition and solving the classical model formulation directly. Presently the algorithms that uses Dantzig-Wolfe given the best results (Desrochers, Desrosiers and Solomon, and Kohl), but the Ph.D. thesis of Kontoravdis shows promising results for using the classical model formulation directly. In this Ph.D. project we have used the Dantzig-Wolfe method. In the Dantzig-Wolfe method the problem is split into two problems: a \master problem" and a \subproblem". The master problem is a relaxed set partitioning v vi problem that guarantees that each customer is visited exactly ones, while the subproblem is a shortest path problem with additional constraints (capacity and time window). Using the master problem the reduced costs are computed for each arc, and these costs are then used in the subproblem in order to generate routes from the depot and back to the depot again. The best (improving) routes are then returned to the master problem and entered into the relaxed set partitioning problem. As the set partitioning problem is relaxed by removing the integer constraints the solution is seldomly integral therefore the Dantzig-Wolfe method is embedded in a separation-based solution-technique. In this Ph.D. project we have been trying to exploit structural properties in order to speed up execution times, and we have been using parallel computers to be able to solve problems faster or solve larger problems. The thesis starts with a review of previous work within the eld of VRPTW both with respect to heuristic solution methods and exact (optimal) methods. Through a series of experimental tests we seek to dene and examine a number of structural characteristics. The rst series of tests examine the use of dividing time windows as the branching principle in the separation-based solution-technique. Instead of using the methods previously described in the literature for dividing a problem into smaller problems we use a methods developed for a variant of the VRPTW. The results are unfortunately not positive. Instead of dividing a problem into two smaller problems and try to solve these we can try to get an integer solution without having to branch. A cut is an inequality that separates the (non-integral) optimal solution from all the integer solutions. By nding and inserting cuts we can try to avoid branching. For the VRPTW Kohl has developed the 2-path cuts. In the separationalgorithm for detecting 2-path cuts a number of test are made. By structuring the order in which we try to generate cuts we achieved very positive results. In the Dantzig-Wolfe process a large number of columns may be generated, but a signicant fraction of the columns introduced will not be interesting with respect to the master problem. It is a priori not possible to determine which columns are attractive and which are not, but if a column does not become part of the basis of the relaxed set partitioning problem we consider it to be of no benet for the solution process. These columns are subsequently removed from the master problem. Experiments demonstrate a signicant cut of the running time. Positive results were also achieved by stopping the route-generation process prematurely in the case of time-consuming shortest path computations. Often this leads to stopping the shortest path subroutine in cases where the information (from the dual variables) leads to \bad" routes. The premature exit from the shortest path subroutine restricts the generation of \bad" routes signi cantly. This produces very good results and has made it possible to solve problem instances not solved to optimality before. The parallel algorithm is based upon the sequential Dantzig-Wolfe based algorithm developed earlier in the project. In an initial (sequential) phase unsolved problems are generated and when there are unsolved problems enough vii to start work on every processor the parallel solution phase is initiated. In the parallel phase each processor runs the sequential algorithm. To get a good workload a strategy based on balancing the load between neighbouring processors is implemented. The resulting algorithm is eÆcient and capable of attaining good speedup values. The loadbalancing strategy shows an even distribution of work among the processors. Due to the large demand for using the IBM SP2 parallel computer at UNIC it has unfortunately not be possible to run as many tests as we would have liked. We have although managed to solve one problem not solved before using our parallel algorithm.

MPI is the standard interface for message passing in parallel systems. The standard was defined in an open collaborative forum involving about 60 people from 40 different organisations, and the finished version was completed in May 1994. EPCC and CRI have established a collaboration to develop and support a robust, high performance implementation of MPI for the T3D. The product was released from EPCC in May 1995. In this paper we describe the architecture of MPI for the T3D and present performance measurements of the implementation. It is shown that a significant part of the T3D capability is available through MPI, demonstrating that MPI is capable of delivering both performance and portability to applications. 1 Introduction MPI [5] is the standard interface for message passing in parallel systems. The standard was defined in an open collaborative forum involving about 60 people from 40 different organisations. An initial draft of the standard was released at the Supercomputing '93 ...

In this study, we are concerned with the parallelizationof finite element mesh generation and its decomposition, and the parallel solution of sparse algebraic equations which are obtained from the parallel discretization of second order elliptic partial differential equations (PDEs) using finite difference and finite element techniques. For this we use the Parallel ELLPACK (//ELLPACK) problem solving environment (PSE) which supports PDE computations on several MIMD platforms. We have considered the ITPACK library of stationary iterative solvers which we have parallelized and integrated into the //ELLPACK PSE. This Parallel ITPACK package has been implemented using the MPI, PVM, PICL, PARMACS, nCUBE Vertex and Intel NX message passing communication libraries. It performs very efficiently on a variety of hardware and communication platforms. To study the efficiency of three MPI library implementations, the performance of the Parallel ITPACK solvers was measured on several distributed memory architectures and on clusters of workstations for a testbed of elliptic boundary value PDE problems. We present a comparison of these MPI library implementationswith PVM and the native communication libraries, based on their performance on these tests. Moreover we have implemented in MPI, a parallel mesh generator that concurrently produces a semi–optimal partitioning of the mesh to support various domain decomposition solution strategies across the above platforms. The results indicate that the MPI overhead varies among the various implementations without significantly affecting the algorithmic speedup even on clusters of workstations.

Users of parallel machines need to have a good grasp for how different communication patterns and styles affect the performance of message-passing applications. LogGP is a simple performance model that reflects the most important parameters required to estimate the communication performance of parallel computers. The message passing interface (MPI) standard provides new opportunities for developing high performance parallel and distributed applications. In this paper, we use LogGP as a conceptual framework for evaluating the performance of MPI communications on three platforms: Cray-Research T3D, Convex Exemplar 1600SP, and a network of workstations (NOW). Our objective is to identify a performance model suitable for MPI performance characterization and to compare the performance of MPI communications on several platforms.

. In this paper we present a portable exact parallel algorithm for the maximum clique problem on general graphs. Computational results with random graphs and some test graphs from applications are presented. The algorithm is parallelized using the Message Passing Interface (MPI) standard. The algorithm is based on the Carraghan-Pardalos exact algorithm (for unweighted graphs) and incorporates a variant of the greedy randomized adaptive search procedure (GRASP) for maximum independent set of Feo, Resende, and Smith (1994) to obtain good starting solutions. 1. INTRODUCTION Let G= (V,E) be an undirected weighted graph where V = {v 1 , v 2 , . . . , v n } is the set of vertices in G, and E #V×V is the set of edges in G. Each vertex v i #V is associated with a positive weight w i . For a subset S #V , we define the weight of S to be W (S) =å i#S w i and G(S) = (S,E #S×S) as the subgraph induced by S. The size of the vertex set is throughout this paper denoted by n. The adjacenc...

Combining parallel programming with prototyping is aimed at alleviating parallel programming by enabling the programmer to make practical experiments with ideas for parallel algorithms at a high level, neglecting low-level considerations of specific parallel architectures in the beginning of program development. Therefore, prototyping parallel algorithms is aimed at bridging the gap between conceptual design of parallel algorithms and practical implementation on specific parallel systems. The essential prototyping activities are programming, evaluation and transformation of prototypes. This paper gives a report on some experience with implementing parallel algorithms based on prototype evaluation and transformation employing the ProSet-Linda approach. 1 Introduction Parallel programming is conceptually harder to undertake and to understand than sequential programming, because a programmer often has to cope with the coexistence and coordination of multiple parallel activities....

PVM and MPI, two systems for programming clusters, are often compared. The comparisons usually start with the unspoken assumption that PVM and MPI represent different solutions to the same problem. In this paper we show that, in fact, the two systems often are solving different problems. In cases where the problems do match but the solutions chosen by PVM and MPI are different, we explain the reasons for the differences. Usually such differences can be traced to explicit differences in the goals of the two systems, their origins, or the relationship between their specifications and their implementations. For example, we show that the requirement for portability and performance across many platforms caused MPI to choose approaches different from those made by PVM, which is able to exploit the similarities of network-connected systems.