| 19 | | * 14h00-15h00 Virginia Niculescu (Babes-Bolya University of Cluj-Napoca), '''On Granularity in Data-Parallel Programs Development'''[[BR]]One key to attaining good parallel performance is choosing the right granularity for the application. Parallel computation models with high level of abstraction, usually, do not have mechanisms for specifying and building granularity. If such mechanisms are introduced they could be very useful since they allow a better evaluation of the performance, and finally, an easier implementation. It is considered that a model of parallel computation, to be useful, must fulfill a set of requirements: abstractness, software development methodology, architecture independence, cost measures, no preferred scale of granularity, and efficiently implementable. The development of the programs correct by construction is also a very important issue in parallel setting. The normal flow in a derivation is to start from a specification, derive and express it using the chosen model, and then adjust it for implementation. The question that could arise is “When should we care about the granularity?” - only at the mapping phase, or starting from the beginning, in the derivation phase. Some case studies will be presented, and their analysis indicate that if we are able to specify and to build granularity from the first levels of design, then the chances to obtain good improvements of the resulted costs increase very much. |
| | 19 | * 14h00-15h15 Virginia Niculescu (Babes-Bolya University of Cluj-Napoca), '''On Granularity in Data-Parallel Programs Development'''[[BR]]One key to attaining good parallel performance is choosing the right granularity for the application. Parallel computation models with high level of abstraction, usually, do not have mechanisms for specifying and building granularity. If such mechanisms are introduced they could be very useful since they allow a better evaluation of the performance, and finally, an easier implementation. It is considered that a model of parallel computation, to be useful, must fulfill a set of requirements: abstractness, software development methodology, architecture independence, cost measures, no preferred scale of granularity, and efficiently implementable. The development of the programs correct by construction is also a very important issue in parallel setting. The normal flow in a derivation is to start from a specification, derive and express it using the chosen model, and then adjust it for implementation. The question that could arise is “When should we care about the granularity?” - only at the mapping phase, or starting from the beginning, in the derivation phase. Some case studies will be presented, and their analysis indicate that if we are able to specify and to build granularity from the first levels of design, then the chances to obtain good improvements of the resulted costs increase very much. |
