Gemini - an object-based Model of Parallel Computation

For many types of scientific computations, clusters of workstations interconnected by a high-speed link are now often considered as a cheaper alternative to distributed-memory computers. The parallel applications for these clusters are mostly programmed using low-level message passing systems such as MPI or PVM, which are general and efficient. However, this efficiency is achieved at the cost of leaving much of the burden of the low-level parallel programming on the application programmer. By restricting the problem domain to that of digital filtering applications, we can provide the programmers with a more abstract model allowing them to express the algorithm more directly. This thesis defines Gemini, an object-based model of parallel computation aimed at processing of large volumes of data on homogeneous clusters of workstations. Gemini provides a generalized data-flow model of parallel computation; the model supports the notion of farms of worker processes interconnected by communication channels and/or shared variables. The workers in farms can be synchronized using locks, barriers and shared coaters. The model supports the creation of prefabricated building blocks; a small set of standard input and output blocks is defined by the model. The model provides a C language interface and can be implemented as a library. To practically validate that the model is logically consistent and can be implemented, its definition was elaborated into an actual shared-memory implementation. An overview of this implementation is presented in the thesis. To build confidence in the implementation is presented in the thesis. To build confidence in the implementation, checks of its conformance with the model were performed using a simple testing framework. The implementation was also extended into a simulator working in model-time which can be used for verification of applications based on the model. Running the applications on this simulator preserves the dynamics of the parallel computation and allows more thorough application testing, e. g. deadlock detection. also included in the thesis is an outline of an initial cluster implementation of the modelAvailable from STL Prague, CZ / NTK - National Technical LibrarySIGLECZCzech Republi