Models of computation for NoC mapping: Timing and energy saving awareness. (February 2017)

• Record Type: Journal Article
• Authors: Marcon, César; Webber, Thais; Susin, Altamiro Amadeu
• Is Part Of: Microelectronics journal. Volume 60(2017)
• Page Start: 129
• Page End: 143
• Publication Date: 2017-02
• Subjects: Application modeling -- Mapping -- Network-on-Chip (NoC) -- Energy minimization -- Performance analysis; Microelectronics -- Periodicals; Microélectronique -- Périodiques; Microelectronics; Electronic journals; Journals - contents and abstracts; Periodicals; 621.3805
• DOI: 10.1016/j.mejo.2016.09.005
• Languages: English
• ISSNs: 0959-8324

Abstract

Abstract: A complex application implemented as a System-on-Chip (SoC) demands extensive system level modeling. Its implementation encompasses a large number of cores and an advanced interconnection scheme such as a Network-on-Chip (NoC). This type of application normally requires energy efficiency and execution time minimization, which implies high-level exploration for cores/tasks placement into the target architecture. A Model of Computation (MoC) captures some characteristics of the applications aiming to fulfill high-level explorations. This work analyzes MoCs employed on the static and dynamic mapping of applications onto regular NoCs, providing a classification based on aspects of computation and communication. Additionally, this paper discusses advantages and drawbacks of these MoCs, such as the complexity of capturing application aspects, as well as the mapping quality. Finally, this work implements the five MoCs more applied on the mapping and compares them applying a benchmark composed of synthetic and embedded applications running on various NoC sizes. Highlights: We propose a classification scheme of applications for task/core mapping onto NoCs. Five graph-based Models of Computation (MoC) are detailed for mapping regular NoCs. MoCs are compared on the easiness of estimating energy and execution time. Mapping simulations with synthetic applications exemplify the modeling effort. New mapping algorithms may benefit from easy capturing of application requirements.