Abstract
Modern heterogeneous multicore systems connected via Network-on-Chip (NoC) demand a sophisticated balance between energy efficiency, real-time performance, and reliability. This paper introduces a joint optimization framework that simultaneously addresses processor frequency scaling (DVFS), selective task duplication, and communication path selection. By formulating the problem as a Mixed-Integer Linear Program (MILP) and providing a scalable three-step heuristic, the proposed method achieves energy savings of up to 39.7% while strictly adhering to reliability and timing deadlines.
Introduction
As chip complexity grows, heterogeneous architectures—combining high-performance and energy-efficient cores—have become standard. However, managing these systems involves navigating three conflicting objectives: minimizing energy consumption via Dynamic Voltage and Frequency Scaling (DVFS), meeting stringent real-time deadlines, and ensuring reliability against soft errors. Lowering voltage to save power often increases task execution time and the probability of transient faults. Existing research frequently treats these issues in isolation or focuses on homogeneous systems. This work bridges the gap with a unified strategy for task placement, fault tolerance, and communication routing in diverse hardware environments.
System Model and Methodology
The target architecture consists of a grid of heterogeneous processing elements ($\theta$) connected through a NoC with routers ($R$), as shown in Figure 1.
The authors formulate the mapping challenge as a Mixed-Integer Nonlinear Program (MINLP). To ensure tractability, they linearize the model into a Mixed-Integer Linear Program (MILP), solvable for small-scale problems. For larger systems, a three-step heuristic is employed:
- Task Allocation & Scheduling: Assigning tasks to cores and determining execution order.
- Frequency Selection & Partial Duplication: Optimizing DVFS levels and identifying "vulnerable" tasks for duplication.
- Multi-path Selection: Choosing optimal NoC routes to balance energy and avoid bottlenecks.
Partial Task Duplication
A core innovation is the use of partial task duplication. Unlike full duplication, which incurs significant energy and communication overhead, this approach only duplicates tasks that cannot meet reliability targets through frequency scaling alone. Figure 2 illustrates various duplication scenarios.
Experimental Evaluation
The framework was evaluated using real-world benchmarks (LU Decomposition, Gaussian Elimination, FFT, and Laplace Equation) and synthetic workloads generated via TGFF.
Key findings include:
- Energy Efficiency: The joint optimization approach reduced energy consumption by up to 39.7% compared to baseline optimal models.
- Scalability: The heuristic produces solutions for complex task graphs in under 1.5 seconds, whereas exact MILP solvers become computationally prohibitive.
- Communication Balance: By utilizing multi-path routing—selecting between time-optimal and energy-optimal paths—maximum processor energy was reduced by approximately 7.5% compared to fixed XY routing.
Conclusion
This research demonstrates that joint optimization of task mapping, frequency scaling, and routing is essential for modern heterogeneous NoCs. By selectively duplicating tasks and intelligently routing communication, systems can achieve significant energy savings without compromising reliability or timing. Future work will explore dynamic scheduling for unpredictable workloads and thermal-aware placement to mitigate chip hotspots.
Summary updated