Semantics and Performance Considerations of Kedama, a GUI-based Massively Parallel Programming Language

Abstract

Kedama is a GUI-based parallel programming language, in which the user can construct and modify the program dynamically even when the program is running. The dynamic nature of the system and its non-technical target audience require different consideration on its design tradeoffs; such as fast and separate compilation of scripts, and simpler control structures for parallel execution.Based on the design choice, the previous version of the Kedama system had provided a usable system. However, the previous semantics model, called "pure-SIMD", had some flaws; when a conditional statement is used, the semantics of script calling was not consistent. Also, the performance was unnecessarily degraded because of the sequentialized execution.This paper presents a new semantics model called "predicated SIMD" that addresses the problems. In the predicated SIMD model, a parallel element (called "a turtle") now has a Boolean value called "predicate" that denotes if the turtle is active or not. In a conditional statement, the predicates for the turtles are set property so that the statements can be executed without being sequentialized.With this improvement, the semantics discrepancy is reduced. Also, there is a factor of 20 improvement in micro benchmarks, and a factor of 6 improvement on the overall performance in real examples.