Vectorized presentation-level services for scientific distributed applications

摘要

The use of heterogeneous distributed systems is a promising approach to significantly increase computational performance of scientific applications. However, one key to this strategy is to minimize the percentage of time spent by an application moving data between machines. This percentage is composed of two parts: (1) the time to translate data between the formats used on different machines, and (2) the time to move data over the network that interconnects the machines. Previous work suggests that data format conversion activity, generally known as presentation-level services, is by far the more costly of the two. In this paper we describe how vectorization can be used to improve presentation-level performance in scientific applications by one or two orders of magnitude over the conventional approach. While others have recognized the advantages of vectorized data format conversion, we describe how to automate this process so that an application programmer need not explicitly call vectorization routines. We explore the impact of presentation-level vectorization on software portability, programming efficiency and protocol standards. We compare our performance results with those of two other popular distributed application programming tools and then summarize the lessons we have learned during the course of our research.