AData ow-basedMassivelyParallelProgrammingLanguage "V" and Its Implementation on A Stock Parallel Machine

We propose a data BLOCKINow-based massively parallel programming language, called \V," which would minimize the diculties in writing massively parallel programs. The language V has both merits of functional programming and object-based programming. Our starting point is a data BLOCKINow-based functional programming language, called \Valid," which we have developed so far, because functional programming paradigm is able to abstract away the timing problem, that is the problem of execution sequence and synchronization control, in writing massively parallel programs. The language V provides an object-based (or multi-agent) abstraction, called \agent," to write parallel entities which have their own states in them and communicate with each other. In addition, we can connect agents explicitly and abstract an ensemble of agents on a predened topology description, called \eld," in order to write a massively parallel program which naturally re BLOCKINects the structure of the problem. As the language V has its basis on a data BLOCKINow-based functional programming language originally designed for data BLOCKINow architecture, it is easy to extract parallelism of various level. In our implementation, the compiler extracts ne-grain parallel threads in the rst stage, and then schedules the ne grain parallel threads and constructs coarser grain threads to exploit as much eective paral-lelism as target machines can provide. In this paper, implementation issues for a distributed-memory parallel machine, Fujitsu AP1000, are discussed. We also show and discuss a preliminary evaluation of our compiler and runtime system developed for AP1000.