Data Flow Pattern Analysis of Scientific Applications

Many parallel and distributed applications have well defined structure which can be described by few simple relations. Some structures are so common that they are abstracted to parallel and grid benchmarks, other structures, are data and application dependent, become evident only at run-time. Many of these patterns reflect flow of data in the applications. Data Flow Graphs (DFG) provide an abstraction that allows to express data flow in the applications and to decompose the application into tasks. Applications represented by task graphs can be translated into MPI or Java RMI parallel programs having a consistent set of Send/Receive calls or set of remote method invocations. In this paper we describe some data flow patterns observed in the Computational Fluid Dynamic (CFD) applications and used in the NAS Parallel Benchmarks and in the NAS Grid Benchmarks. Then we describe new approach to the data flow analysis of the traces of parallel programs via the Data Cube Operator. The approach uses On-Line Analytic Processing (OLAP) to create various views of application trace files. The detected patterns can help in application performance tuning and improving compiler support for the commonly arising patterns. Finally, we show examples of using DFGs in parallel and distributed scientific programming.