DPS: A Distributed Program Simulator and Performance Measurement Tool

The recent past has seen an intense e ort to harness the power of networked workstations as an alternative to expensive supercomputers. However, the e cacy of parallel programming in a distributed system is limited by several factors. These include (1) the nature of the problem being attacked, (2) the design of the application program, (3) the speeds of the various components of the network, and (4) the communication between nodes in the system. A program that performs well in one particular environment may be extremely ine cient in another. The design of an e cient parallel program generally relies on detailed knowledge about the network hardware on which the program will run. In the design phase of program development, issues such as the programming paradigm, the granularity of parallelism, and the cost of intertask communication must be addressed. Programmers must predict how a particular design will perform in practice, but complex program behavior makes it di cult to determine how characteristics of a particular network will a ect each issue. Therefore, even careful consideration for hardware is insu cient to guarantee an e cient implementation. DPS (Distributed Program Simulator) helps programmers build e cient parallel programs by alleviating the di culty of matching applications to hardware. The user provides DPS with a program blueprint that describes the number of tasks, the location of tasks, and the characteristics of intertask communication within a parallel program. The method for specifying blueprints is simple and general. Blueprints for many di erent classes of parallel programs can be described without the complexity of a full implementation. DPS uses the information in the blueprint to run a real-time simulation of the program on the network for which it is intended. Upon completion of the simulation, DPS presents the user with a detailed performance pro le. The pro le can be used to identify bottlenecks and tune the algorithm. In this way, the programmer can develop a program that best matches the network hardware. To simplify blueprint construction, DPS introduces a model that represents an application by its shape, terrain, and texture. These properties characterize the program based on (1) the decomposition of the program into functional units, (2) the interaction between fuctional units, (3) the grouping of functional units into processes, and (4) the mapping of processes onto machines in