Invasive Computing for Mapping Parallel Programs to Many-Core Architectures by Andreas Weichslgartner & Stefan Wildermann & Michael Glaß & Jürgen Teich

Author:Andreas Weichslgartner & Stefan Wildermann & Michael Glaß & Jürgen Teich
Language: eng
Format: epub
ISBN: 9789811073564
Publisher: Springer Singapore

4.7 Summary

In this chapter, we have described a novel and distributed methodology to map applications, specified by application graphs, onto NoCs. We refer this methodology as self-embedding. The core idea of this algorithm is that each task is responsible for mapping its succeeding tasks and messages. To prove the power of decentralized self-embedding, we have presented a model and interface for this class of algorithms and also have introduced two incarnations. We showed that a local best neighbor (BN) algorithm can compete with the global path load (PL) algorithm by offering a magnitude less monitor overhead. Also, we have developed an improved random-walk-based algorithm that offers several advantages over a random-walk algorithm using an equal distribution. In addition, we presented a concept and implementation of a hardware-based acceleration of the self-embedding algorithm which may speedup the mapping process. Dedicated hardware modules, so-called self-embedding modules (SEMs), inside NoC routers perform the distributed application mapping based upon a serialized application graph. The application graph can be split up and processed in a streaming-based manner. This mechanism can be used to accelerate the invasion process and, hence, increases the invasive speedup and efficiency according to Sect. 2.​3.​1. Together with hardware extensions which accelerate the infection, such as the Cicic (see Sect. 2.​4.​3) and the i-NA with task spawning support [17], the invasive overheads can be kept minimal.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.