Distributed Memory Parallel Numerical Computation Communication Library System

Message passing programming is main paradigm in distributed memory systems with which high efficient and scalable parallel programs can be designed. Currently, common message passing systems are PVM and MPI etc., however, they are general libraries for general problems, and thus their application program interfaces (API) are not as easily usable for numeral computing programming as we would like. In order to develop rapidly parallel numerical software with high performance and reliability, it is much necessary to research and develop communication libraries for a certain numerical computing field. This paper gives an in-depth research of communication library systems for numerical computing from application program interfaces, collective communications, and testing. Main contributions achieved in this paper are as follows:First, based on analysis of current communication library systems and features of numerical computing, it is pointed out that design of application program interface (API) is important in design of communication libraries for numerical computing. It is concluded that the application program interface can be improved from following three aspects, and some examples are given to illustrate how to realize them: (1) abstract name of processes; (2) selectivity of message based on source; (3) non-blocking communication.Second, formally give a definition of logical topology for one-to-all broadcasts and its fundamental features. Propose an algorithm of one-to-all broadcasts based on this formal definition. Here logic topology means a kind of mechanisms that determine how the data in a distributed operation are sent. Due to the abstraction of logic topology, the algorithm is a framework for broadcast algorithms. When a concrete logic topology is given, an algorithm of broadcast based on it can be derived from the framework of algorithm. Therefore, this gives a new solution to design of broadcast algorithms.Third, propose a formal definition of logic topology for all-to-one reductions, and obtain its basic features. An algorithm of all-to-one reduction based on the formal definition is given. Because of abstraction of logic topology, the algorithm is a framework of all-to-one reduction algorithms. When a concrete logic topology is given, an algorithm of all-to-one reduction can be obtained from the framework. Thus, this supplies a new approach to designing algorithms of all-to-one reductions.