Retiming and recycling are two transformations used to optimize the performance of latency-insensitive (a.k.a. synchronous elastic) systems. This paper presents an approach that combines these two transformations for performance optimization of elastic systems with early evaluation. The method is based on Mixed Integer Linear Programming.On a set of random benchmarks the proposed method achieves, in average, 14.5% performance improvement over min-delay retiming configurations.

Abstract. Digital electronic systems typically use synchronous clocks and primarily assume fixed duration of their operations to simplify the design process. Time elastic systems can be constructed either by replacing the clock with communication handshakes (asynchronous version) or by augmenting the clock with a synchronous version of a handshake (synchronous version). Time elastic systems can tolerate static and dynamic changes in delays (asynchronous case) or latencies (synchronous case) of operations that can be used for modularity, ease of reuse and better power-delay trade-off. This paper describes methods for the modeling, performance analysis and optimization of elastic systems using Marked Graphs and their extensions capable of describing behavior with early evaluation. The paper uses synchronous elastic systems (aka latency-tolerant systems) for illustrating the use of Petri Nets, however most of the methods can be applied without changes (except changing the delay model associated with events of the system) to asynchronous elastic systems.