Rapid design space exploration of heterogeneous embedded systems using symbolic search and multi-granular simulation

  • Authors:

  • S. Mohanty;V. K. Prasanna;S. Neema;J. Davis

  • Affiliations:

  • University of Southern California, Los Angeles, CA;University of Southern California, Los Angeles, CA;Vanderbilt University, Nashville, TN;Vanderbilt University, Nashville, TN

  • Venue:
  • Proceedings of the joint conference on Languages, compilers and tools for embedded systems: software and compilers for embedded systems
  • Year:
  • 2002

Quantified Score

Hi-index 0.00

Visualization

Abstract

In addition to integrating different Intellectual Property cores, heterogeneous embedded systems provide several architecture knobs such as voltage, operating frequency, configuration, etc. that can be varied to optimize performance. Such flexibilities results in a large design space making system optimization a very challenging task. Moreover, such systems operate in mobile and other power constrained environments. Therefore, in addition to rapid exploration of a large design space a designer has to optimize both time and energy performance. To address these issues, we propose a hierarchical design space exploration methodology. Our methodology initially uses symbolic constraint satisfaction to rapidly prune the design space. This pruning process is followed by a system wide performance estimation to further reduce the number of candidate designs. Finally, detailed simulation using low-level simulators are performed to select an appropriate design. Our methodology is implemented by integrating two tools, DESERT and HiPerE, into the M model based Integrated simuLAtioN (MILAN) framework. DESERT uses Ordered Binary Decision Diagrams based symbolic search to rapidly explore a large design space and identifies candidate designs that meet the user specified performance constraints. HiPerE provides rapid estimation of system wide energy and latency based on component level simulations and also facilitates energy optimization. MILAN provides the required modeling support for these tools and also facilitates component specific multi-granular simulations through seamless integration of various simulators.