Aggregate Update Problem for Multi-clocked Dataflow Languages

Abstract

Dataflow languages have, as well as functional languages, immutable semantics, which is often implemented by copying values. A common compiler optimization known from functional languages involves analyzing which data structures can be modified in-place instead of copying them. This paper presents a novel algorithm to this so called Aggregate Update Problem for multi-clocked dataflow languages, i.e. those that allow streams to have events at disjoint timestamps, like e.g. Lucid, Lustre and Signal. Unrestricted multi-clocked languages require a static triggering analysis on how events and hence data values are read, written and replicated. We use TeSSLa as a generic stream transformation language with a small set of operators to develop our ideas. We implemented the solution in a TeSSLa compiler targeting the Java VM via Scala code generation which combines persistent data structures and mutable data structures for those data values which allow in-place editing. Our empirical evaluation shows considerable speedup for use cases where queues, maps or sets are dominant data structures.