Himomodel.si

Task 3.2: Parallelization

Identified bottlenecks will be examined in detail, and it will be determined whether the parallelization of a particular bottleneck would contribute to the overall algorithm speedup. In case of the positive assessment, we will decide on the most appropriate parallelization strategy. It is expected, that low­level computations will be best suited for fine­grained massively parallel execution on the graphic cards (GPUs) and that higher­level computations will benefit more from coarse­grained functional decomposition parallelization strategies, which would allow them to run in parallel on multiple cores (CPUs) of computing nodes. Finally, algorithm in question will be ported or redesigned for parallel processing. It is our opinion that the hierarchical compositional structure offers excellent opportunities for efficient parallelization. In addition to implementation of low­level algorithms on massively parallel architectures, there is also an opportunity to exploit inherently parallel structure in the upper layers of compositional model. We plan to introduce speculative computing on the potentially idling resources as well. The reasoning through the layers of the model will be done by evaluating hypotheses, and selecting the most probable ones. The algorithm traverses through the hierarchy, its direction is determined by the hypothesis evaluation. If it arrives into a dead end, a part of its path through the hierarchy has to be re­evaluated. This offers exciting possibility of speculative parallelization, where additional computing resources, such as idle processor cores, perform evaluation of less likely hypotheses, in hope that they might be useful in future processing steps, and thus minimizing the potential impact of suboptimal decision at any of the stages.

• < Prev