Takayuki Nakatomi

Development of General-Purpose Processing Element and Network-Based Dataflow Processing System: Part One

This paper introduces a system, called Ubiquitous Multi-Processor (UMP) system, which constructs a dataflow network dynamically by connecting function-specific execution modules. We developed a general-purpose execution module (GP-PE) which performs an equivalent function to SS-PE by downloading the software library to improve the efficiency of development and reconfigure the function dynamically. This paper shows overhead time is 200ms caused by network communication among PEs using 1000 Base-T, on decentralized network where a client module, management modules, and calculation modules including GP-PE are deployed, comparing it with the system performance on a single PC. The experimental result presents the adaptability of GP-PE to UMP system in general use. Moreover, the ratio of the overhead time of the library to downloading time for general using is clarified. It becomes useful parameter to decide method of downloading DPL.

Network-Based Data Flow Processing System with Variable Data Granularity for Inter-module Communication

Integration of many functions in mobile devices often results in the insufficiency of computing resources and security problems. We utilize multiple computing modules distributed in the environment as external resources, and call this processor network system “UMP (Ubiquitous Multi Processor) system.” In our previous research, the basic architecture of a UMP system and its communication theory were verified. This paper evaluates the correlation between data granularity and system performance for applications with variable data granularity and verifies the usefulness of these parameters in dynamic scheduling necessary to construct intelligent systems. A JPEG encoding application was evaluated on the UMP system with six sub-functional modules. Optimization of data granularity shows 180% improvement of system performance. The experimental result shows that data granularity of inter-PE communication could be an important parameter in optimizing the system architecture and dynamic scheduling. Moreover, the result shows the possibility of further performance improvement through optimization of data granularity by focusing on individual modules.