Deterministic memory sharing in Kahn process networks: Ultrasound imaging as a case study

Abstract

Kahn process networks are a popular programming model for programming multi-core systems. They ensure determinacy of applications by restricting processes to separate memory regions, only allowing communication over FIFO channels. However, many modern multi-core platforms concentrate on shared memory as a means of communication and data exchange. In this work, we present a concept for deterministic memory sharing in Kahn process networks. It allows to take advantage of shared memory data exchange mechanisms on such platforms while still preserving determinacy. We show how any Kahn process network can be transformed to use deterministic memory sharing by giving a set of transformations that can be applied selectively, only looking at one process at a time. We demonstrate how these techniques can be applied to an ultrasound image reconstruction algorithm. For an implementation on a test system, our technique yields significantly better performance combined with a drastically smaller memory footprint.

Minimising Access Conflicts on Shared Multi-Bank Memory

“…Rihani et al [28] propose a solution for single-producer single-consumer process networks by assigning each consumer a bank for all its input FIFOs. This works well for the selected application model, which, however, necessitates large amounts of data copying and is therefore inherently inefficient on shared memory platforms [35]. Finally, Goens et al [13] employ a buffer allocation approach similar to this paper.…”

“…The model is a simplified representation of the OpenVX execution model [11] and the Memory Interference Graph [12], but it is intentionally generic to be applicable to a broad class of different execution models, for instance mechanisms like Deterministic Memory Sharing in Kahn Process Networks [35] or traditional custom multi-threading with manual thread synchronisation.…”