This paper presents a common data model for both shared and distributed parallel programming. Our model aims at parallel application programming, especially in the field of data parallelism. We introduce coordinators - shared objects which automatically perform synchronization in shared memory programming and remote communication in distributed programming. Programming with coordinators allows a static definition of dynamic access properties (shared access patterns) in a declarative manner. Thus, there is no need for explicit use of synchronization or communication primitives in order to get well defined and efficient parallel programs. In contrast, declarativity allows to think in abstract coordination categories. Our paradigm is named Declarative Imperative Parallel Programming due to this fact and due to our commitment to imperative applications, mainly from the field of scientific computing, image processing, simulation and optimization. 1. The DIPP programming model Ideally, ex...

. This paper presents an architecture independent paradigm for explicit parallel programming. The focus is on the coordination (data) model: we introduce coordinators, an uniform passive object approach for distributed and shared variables programming. Programming with coordinators allows static definition of dynamic access properties (access patterns) in a declarative manner. These access specifications lead to automatic synchronization or remote communication, depending on the underlying architecture. Coordinators are very well suited for generalized data parallel programming thus enabling comfortable and efficient programming in application fields like scientific computing and image processing. Nevertheless they can be used for task parallelism, too. While the distributed implementation is currently being designed the shared memory implementation already shows good results. 1 Overview In conventional explicit parallel programming there are two aspects to be considered:...

We present a programming model which offers ease of parallel application programming for both shared variables and distributed programming. The model is more abstract than distributed models such as PVM, MPI and Split C and shared models such as thread packages or the ANL Parmacs. It improves on the property of memory-modelindependence of other models such as CC++ and Nexus and offers a remarkable amount of ease of use. ffl The process model - a work pool - is akin to CORBA in that it is defined in terms of abstract work pieces (services, tasks) rather than processes or threads. Semantically, this corresponds to asynchronous remote procedure calls. ffl The data model features global data which allows memory-model-independence as well. The coordination of accesses to global data is possible by so-called coordinators. Coordinators enable the declarative specification of access patterns to global data thus allowing automatic coordination. The overall model allows considering ...

This paper introduces a problem solving method involving independent agents and a set of partial solutions. In the Iterative Multi-Agent (IMA) method, each agent knows about a subset of the whole problem and can not solve it all by itself. An agent picks a partial solution from the set and then applies its knowledge of the problem to bring that partial solution closer to a total solution. This implies that the problem should be composed of subproblems, which can be attended to and solved independently. When a realworld problem shows these characteristics, then the design and implementation of an application to solve it using this method is straightforward. The solution to each sub-problem can affect the solutions to other sub-problems, and make them invalid or undesirable in some way, so the agents keep checking a partial solution even if they have already worked on it. The paper gives an example of constraint satisfaction problem solving, and shows that the IMA method is highly parallel and is able to tolerate hardware and software faults.