Search-space Decomposition for System-level Design Space Exploration of Embedded Systems

Abstract

The development of large-scale multi- and many-core platforms and the rising complexity of embedded applications have led to a significant increase in the number of implementation possibilities for a single application. Furthermore, rising demands on safe, energy-efficient, or real-time capable application execution make the problem of determining feasible implementations that are optimal with respect to such design objectives even more of a challenge. State-of-the-art Design Space Exploration (DSE) techniques for this problem demonstrably suffer from the vast and sparse search spaces posed by modern embedded systems, emphasizing the need for novel design methodologies in this field. Based on the idea of reducing problem complexity by a suitable decomposition of the system specification—in particular, by a reduction of target architecture or task mapping options—the work at hand proposes a portfolio of dynamic decomposition mechanisms that automatically decompose any system specification based on a short pre-exploration of the complete system. We present a two-phase approach consisting of (a) a set of novel data extraction and representation techniques combined with (b) a selection of filtering operations that automatically extract a decomposed system specification based on information gathered during pre-exploration. In particular, we employ heat map data structures and threshold as well as graph-partitioning filters to reduce problem complexity. The proposed decomposition procedure can seamlessly be integrated in any DSE flow, constituting a flexible extension for existing DSE approaches. Furthermore, it improves existing static decomposition techniques and other heuristics relying on information about the problem instance, since systems with irregular architectural topology or distribution of resource types can now be decomposed based on an automatic, problem-independent pre-exploration phase. We illustrate the efficiency of the proposed decomposition portfolio applied to state-of-the-art DSEs for many-core systems as well as networked embedded systems from the automotive domain. Experimental results show significant increases in optimization quality of up to 87% within constant DSE time compared to existing approaches.

Embeddings of Task Mappings to Multicore Systems

“…The work from Richthammer and others [26,27,25] is very similar in nature to the applications discussed in this paper. They also aim to improve Design-Space Exploration (DSE) methods by statically exploiting the architectural structure, although the concrete structure they exploit is different.…”

Generative Design of the Architecture Platform in Multiprocessor System Design

“…An example of non-exact DSE at the ESL can be found in [21], where the authors used genetic algorithms to map tasks on a fixed architecture template of eight homogeneous processors connected via a single shared bus. Similarly, Richthammer et al [6,22] proposed strategies for improving the efficiency of meta-heuristic DSE approaches, again mapping tasks to resources in fixed architecture templates.…”

“…Many previous state-of-the-art DSE publications [3,[5][6][7][8][9][10][11] are based on or are similar in nature to the system model originally presented by Blickle et al in [24], which uses fixed architecture graphs. They all have in common that they map tasks directly to processing elements in a fixed architecture graph.…”

“…However, most traditional approaches are limited because they only consider a fixed architecture template with a fixed set of available hardware components [3,[5][6][7][8][9][10][11]. Alterations in the architecture between design points are only achieved by not allocating idle hardware components.…”

Hybrid Application Mapping for Composable Many-Core Systems: Overview and Future Perspective

“…The authors of Reference [64] propose a second possibility of architecture decomposition that is better suited for irregular architectural topologies or for cases where an a-priori decision about the number and type of resource instances to be removed cannot be made. There, a dynamic decomposition approach is presented which utilizes information from a short preliminary DSE, that is, a pre-exploration based on the complete architecture, to determine resources to be pruned dynamically for the actual extensive DSE.…”