Abstract. The trend from single processor to parallel computer architectures has increased the importance of parallel computing. To support parallel computing it is important to map parallel algorithms to a computing platform that consists of multiple parallel processing nodes. In general different alternative mappings can be defined that perform differently with respect to the quality requirements for power consumption, efficiency and memory usage. The mapping process can be carried out manually for platforms with a limited number of processing nodes. However, for exascale computing in which hundreds of thousands of processing nodes are applied, the mapping process soon becomes intractable. To assist the parallel computing engineer we provide a model-driven approach to analyze, model, and select feasible mappings. We describe the developed toolset that implements the corresponding approach together with the required metamodels and model transformations. We illustrate our approach for the well-known complete exchange algorithm in parallel computing.

Collective Communication Algorithms for 2D torus networks have been investigated quite extensively in the literature and two broad approaches, namely direct methods and indirect (message combining) methods are recognized in the field. While direct methods minimize the volume of data, the indirect methods reduce the number of message start-ups. Consequently, either a suite of algorithms must be employed for efficiency over a wide range of message lengths and communication operations or algorithms should be able to adapt themselves to the current case, possibly by switching between direct and indirect routing modes as appropriate. In this paper, we propose adaptive routing algorithms for all-port, wormhole routed, synchronous, 2D torus networks optimized for one-to-all broadcast, gossiping and complete exchange collective communication operations. The proposed algorithms employ completely-connected subnetworks where complete exchange amongst the nodes in the subnetwork can be accomplished in one step only. Combined with suitable 2D plane tiling techniques, the proposed algorithms share the same set of primitive operations and yield superior performance compared to previously proposed methods, either pure or hybridized.

Parallel and distributed simulations (PADS) realize the distributed execution of a simulation system over multiple physical resources. To realize the execution of PADS, different simulation infrastructures such as HLA, DIS and TENA have been defined. Recently, the Distributed Simulation Engineering and Execution Process (DSEEP) that supports the mapping of the simulations on the infrastructures has been defined. An important recommended task in DSEEP is the evaluation of the performance of the simulation systems at the design phase. In general, the performance of a simulation is largely influenced by the allocation of member applications to the resources. Usually, the deployment of the applications to the resources can be done in many different ways. DSEEP does not provide a concrete approach for evaluating the deployment alternatives. Moreover, current approaches that can be used for realizing various DSEEP activities do not yet provide adequate support for this purpose. We provide a concrete approach for deriving feasible deployment alternatives based on the simulation system and the available resources. In the approach, first the simulation components and the resources are designed. The design is used to define alternative execution configurations, and based on the design and the execution configuration; a feasible deployment alternative can be algorithmically derived. Tool support is developed for the simulation design, the execution configuration definition and the automatic generation of feasible deployment alternatives. The approach has been applied within a large-scale industrial case study for simulating Electronic Warfare systems.

The need for high-performance computing together with the increasing trend from single processor to parallel computer architectures has leveraged the adoption of parallel computing. To benefit from parallel computing power, usually parallel algorithms are defined that can be mapped and executed on parallel computing platforms. In general, different alternative mappings can be defined each with different performance. For small computing platforms with a limited number of processing nodes, the mapping process can be carried out manually. However, for large-scale parallel computing platforms in which hundreds of thousands of processing nodes are applied, the number of possible mapping alternatives increases dramatically, and the mapping process becomes intractable for the human engineer. To assist the parallel computing engineer, we provide a systematic approach to derive feasible mapping alternatives of parallel algorithms to parallel computing platforms. The approach includes activities for modeling the parallel algorithm and parallel computing platform, generation of feasible mapping alternatives, generation of the deployment code, and finally the deployment of the generated code to the nodes. We evaluate our approach for deriving feasible mapping alternatives for four well-known parallel algorithms. The evaluation is based on both simulations and real executions of the generated mapping alternatives.

No abstract