Multi rectangle modeling approach for application mapping on a many-core system




Igor Tcarenko, Mohammad Fattah, Pasi Liljeberg, Juha Plosila, Hannu Tenhunen

Marco Aldinucci, Daniele D’Agostino, and Peter Kilpatrick

Euromicro international conference on parallel, distributed and network-based processing

PublisherIEEE Computer Society

2014

2014 22nd Euromicro International Conference on Parallel, Distributed, and Network-Based Processing

Proceedings - 2014 22nd Euromicro International Conference on Parallel, Distributed, and Network-Based Processing, PDP 2014

452

457

6

978-1-4799-2728-9

1066-6192

DOIhttps://doi.org/10.1109/PDP.2014.99

http://api.elsevier.com/content/abstract/scopus_id:84899433072



The importance of first node selection in run-time resource management is shown in our previous work, SHiC. It is desired in SHiC to find the optimum node in an agile manner. Accordingly, the current mapping picture of the system is simplified to SHiC by modeling each application as a rectangle of occupied nodes. However, the algorithm performance can be influenced significantly with the rectangle model of each application. In this work, we introduce a precise description of our new accurate rectangle modeling algorithm. Moreover, we show that it is not sufficient to always model an application with only one rectangle, as dispersion of the allocated nodes is an irrepressible phenomenon. Accordingly, our algorithm enables to model a mapped application with several rectangles by tuning the model accuracy against the algorithm complexity. However, the algorithm is not in the critical path of the applications executions. Our results shows up to 5% reduction in power dissipation of the network. © 2014 IEEE.




Last updated on 2024-26-11 at 14:13