Position Paper: Logic Programming for Parallel Irregular Applications

Abstract

Parallel computers are becoming increasingly heterogeneous and correspondingly more difficult to program as a result. Irregular applications exacerbate this problem further, given that this class of applications is more diverse and uses different performance features of a computer system than more common application classes. Thus, new approaches are necessary to achieve performance and productivity simultaneously. Domain-specific languages are becoming increasingly popular for high-performance computing, both in the domains of regular (e.g., SPIRAL) and irregular (e.g., Green-Marl) applications. However, past languages for graph computations tend to be too limited to efficiently express the wide range of irregular algorithms needed in applications. Instead, this position paper advocates adapting a language from the database community, Datalog, to the domain of high-performance irregular applications. Although the plain Datalog language is also insufficient for the class of applications targeted, extensions can be added to increase its expressiveness. Starting with a standard language also enables taking advantage of the literature on Datalog implementations, including in the contexts of parallelism and incremental execution of algorithms. Thus, this approach promises to be a good way to implement irregular applications with both productivity and performance.