Advancement of parallel systems skeletons

点击翻译按钮获取中文摘要

The idea of design patterns has been widely researched and applied in parallel application design and development. In most existing works on patterns in parallel programming, the emphasis is on aiding an application developer during the design and development phases of an application. In addition to the traditional applications of patterns, this proposed research extends the use of patterns towards (i) systems, dependability and performance specific aspects of parallel programs, targeting two categories of users: application and system designers/developers; and (ii) assisting in non-trivial, practical, and useful parallel application domains. More specifically, the proposed research focuses on the following systems and dependability specific aspects: (a) fault-tolerance: fault-tolerance strategies differ for different patterns in order to minimize overhead. For example, fault-tolerance strategy for a master-worker pattern with no interaction among workers will differ from the scenario where workers form distinct communication subgroups. This necessitates proper design of strategies specific to patterns; (b) atomicity: since patterns abstract the behavioral requirements of an application, atomicity requirement checking of parallel programs could be facilitated by the use of patterns; this needs extensive research. From the performance side: (c) scheduling strategies for different patterns differ on a heterogeneous computing environment in order to minimize idling time, reduce communication cost and thus enhance efficiency. For example, a task farm with independent dynamic tasks needs to be scheduled differently from a pipeline or a systolic array. Hence designing of appropriate scheduling strategies is crucial for performance. In the non-trivial applications category: (d) two such application domains are recently being investigated and will be reported in the main proposal. In all of these works, the targeted hardware platform is a commodity heterogeneous parallel computing environment, consisting of nodes with multi-core CPUs and GPUs of dissimilar architectures interconnected via high-speed network, where achieving efficiency through proper utilization of resources is a challenge and dependability is a requirement for the developed application.