Programming and Runtime Supports for SPMD Cluster Computing

this paper, we describe the programming environment called DAME (DAta Migration Environment) that supports the development and execution of SPMD applications on heterogeneous clusters. DAME overcomes the difficulties of SPMD cluster computing by means of transparent supports that hide irregular network topology, dynamically adapt data distribution to platform conditions, and mask the consequent non-uniform distribution to the programmer. The originality of DAME with respect to frameworks that share analogous goals are the following: - DAME adopts a traditional programming paradigm such as SPMD, while other tools are oriented to different models. For example, Dataparallel C supports SIMD computations [Ned93], Piranha supports Linda programs [Car95]. - DAME provides data migrations, while other frameworks such as Utopia [Zho93], UPVM and MPVM [Cas94, Cas95] are oriented to task migrations. - DAME includes applications with explicit message-passing while other tools that provide data migration (such as ADM [Cas94] and the framework described in [Ham95]) are oriented to master-slave applications without inter-task communications. - DAME is highly portable because is built on PVM. Moreover, it provides transparent supports for static and dynamic load balancing.