TL;DR
As we move toward many-core processors, traditional electrical interconnects are becoming a major bottleneck. This paper proposes a novel admission control mechanism for on-chip nanophotonic crossbars that uses convex optimization to fairly and efficiently allocate wavelengths and buffer space, resulting in up to three orders of magnitude lower latency in congested "hot spot" traffic scenarios.
Introduction: The Interconnect Bottleneck
In the "billion-transistor era," the way cores communicate on a chip (Network-on-Chip or NoC) is just as important as the cores themselves. Traditional copper-wire interconnects face a "scalability wall": they consume too much power, generate excessive heat, and suffer from high latency as the number of cores increases.
Nanophotonics—using light instead of electricity to move data—offers a high-bandwidth, low-latency alternative. However, optical crossbars provide a massive pool of shared resources, including wavelengths and buffers.
Without a "traffic cop" to manage these resources, a few greedy cores can starve others, or the system can become inefficient during traffic bursts. The authors tackle the challenge of creating an admission control system that is both fair (ensuring all cores get their turn) and efficient (maximizing the chip's total data throughput).
Methodology: Network Utility Maximization
The authors frame resource management as a Network Utility Maximization (NUM) problem. They assign a "utility function" to each connection, using $\alpha$-fair utility functions to mathematically tune the balance between total throughput and fairness.
This convex optimization problem is solved using a diagonally scaled gradient projection algorithm. Designed to be computationally lean and highly scalable, this approach is suitable for implementation in a central on-chip controller. For sudden traffic spikes where iterative solvers might be too slow, the authors provide a fast, non-iterative "closed-form" approximation specifically for bursty conditions.
Architectural Design and Resource Pooling
A key insight of this work is Total Resource Pooling. Unlike previous architectures that often dedicate specific waveguides to specific receivers, this approach treats the entire set of wavelengths as a global pool. This flexibility is what allows it to handle "Hot Spot" traffic significantly better than static designs.
The paper proposes two physical hardware variants for the controller: a low-latency version (prioritizing speed) and a low-power version (prioritizing energy efficiency). Despite being a centralized "traffic cop," the controller is remarkably efficient; it occupies only 1.5% of the total chip area in the low-latency configuration and a negligible 0.2% in the low-power version.
Experimental Evaluation and Results
The system (dubbed MWMR-AC) was evaluated using synthetic traffic patterns (Uniform and Hot Spot) and real-world traces from the SPLASH-2 and PARSEC suites injected into the Sniper cycle-accurate simulator.
- Massive Latency Reduction: In "Hot Spot" traffic, MWMR-AC achieved an average latency three orders of magnitude lower than the baseline Corona architecture.
- Higher Throughput Ceiling: The system hits a saturation throughput of approximately 0.7 (normalized), whereas the baseline saturates at around 0.4–0.5.
- Service Differentiation: By adjusting weights in the utility function, the authors demonstrated precise control over bandwidth allocation for different application classes, enabling true Quality of Service (QoS) on-chip.
- Application Performance: For communication-heavy benchmarks like bodytrack and dedup, the proposed mechanism reduced packet latency by 4.7x and 6.04x, respectively.
Conclusion and Future Work
While the centralized controller introduces a slight overhead in very low-traffic scenarios, its ability to manage congestion makes it a critical component for high-performance many-core systems. Future work may explore decentralized versions of the algorithm or tighter integration with OS-level task schedulers to further optimize system-wide energy consumption.
Summary updated