Tupleware: a distributed tuple space for the development and execution of array-based applications in a cluster computing environment

This thesis describes Tupleware, an implementation of a distributed tuple space which acts as a scalable and efficient cluster middleware for computationally intensive numerical and scientific applications. Tupleware is based on the Linda coordination language (Gelernter 1985), and incorporates additional techniques such as peer-to-peer communications and exploitation of data locality in order to address problems such as scalability and performance, which are commonly encountered by traditional centralised tuple space implementations. Tupleware is implemented in such as way that, whilepr ocessing is taking place, all communication between cluster nodes is decentralised in a peer-to-peer fashion. Communication events are initiated by a node requesting a tuple which is located on a remote node, and in order to make tuple retrieval as efficient as possible, a tuple search algorithm is used to minimise the number of communication instances required to retrieve a remote tuple. This algorithm is based on the locality of a remote tuple and the success of previous remote tuple requests. As Tupleware is targetted at numerical applications which generally involve the partitioning and processing of 1-D or 2-D arrays, the locality of a remote tuple can generally be determined as being located on one of a small number nodes which are processing neighbouring partitions of the array. Furthermore, unlike some other distributed tuple space implementations, Tupleware does not burden the programmer with any additional complexity due to this distribution. At the application level, the Tupleware middleware behaves exactly like a centralised tuple space, and provides much greater flexibility with regards to where components of a system are executed. The design and implementation of Tupleware is described, and placed in the context of other distributed tuple space implementations, along with the specific requirements of the applications that the system caters for. Finally, Tupleware is evaluated using several numerical and/or scientific applications, which show it to provide a sufficient level of scalability for a broad range tasks. The main contribution of this work is the identification of techniques which enable a tuple space to be efficiently and transparently distributed across the nodes in a cluster. Central to this is the use of an algorithm for tuple retrieval which minimises the number of communication instances which occur during system execution. Distribution transparency is ensured by the provision of a simple interface to the underlying system, so that the distributed tuple space appears to the programmer as a single unified resource. It is hoped that this research in some way furthers the adoption of the tuple space programming model for distributed computing, by enhancing its ability to provide improved performance, scalability, flexibility and simplicity for a range of applications not traditionally suited to tuple space based systems.