TL;DR
This paper introduces a predictive power management scheme for multicore communication processors that significantly reduces energy consumption by proactively adjusting core activity and frequency based on predicted traffic loads. By outperforming reactive schemes, it offers up to 40% more power savings, crucial for the ever-growing energy demands of network infrastructure.
What Problem Does This Paper Address?
The internet's energy consumption is a growing concern, with routers being major contributors. Modern routers rely on powerful multicore communication processors designed to handle peak traffic loads, which occur rarely. This over-provisioning leads to significant energy waste during periods of lower traffic. Existing power management techniques are often too conservative or reactive, meaning they only adjust power after a load change, leading to a lag and missed opportunities for savings. Furthermore, general-purpose processor power management policies are not well-suited for the unique traffic patterns and demands of network processors. This paper aims to bridge this gap by proposing a predictive, traffic-aware power management strategy specifically for these processors.
Methods
The core of the proposed solution is a Predictive Traffic Aware Power Management (PTM) scheme. PTM makes three key decisions:
- Predicting Load: It uses Double Exponential Smoothing (DES) for traffic prediction. DES is chosen for its low overhead and accuracy in handling data with trends, making it suitable for network traffic patterns.
- Deciding Number of Active Cores: PTM introduces a novel parameter called Traffic_Factor. This factor combines predicted traffic arrival rates (PPS) with application processing requirements (Cycles Per Packet - CPP) to estimate the actual processing load. This is a crucial improvement over simply using packet rates, as different applications have varying computational needs. The Traffic_Factor is then used to determine the optimal number of active cores (
N_opt). - Deciding Core Frequency: For the active cores, PTM employs Dynamic Voltage and Frequency Scaling (DVFS). To decide the appropriate frequency, it monitors the input queue length. A low-pass filter (exponential smoothing) is used to calculate the average queue length, smoothing out short-term traffic bursts. If the queue is nearly empty, frequencies are lowered; if it starts to fill, frequencies are increased. This approach allows for fine-grained power adjustments at a faster timescale than core activation/deactivation.
The scheme also incorporates C-states (for idle power) and P-states (for active power) to maximize energy savings.
Datasets
The paper uses real-world network traces from CAIDA's equinix-chicago and equinix-sanjose monitors. These traces are one hour long and captured in 2011. They also use synthetic traces to evaluate performance at different traffic rates.
Key Insights and Ideas
- Predictive vs. Reactive: Proactive power management, by predicting traffic and load, significantly reduces the lag between load changes and power adaptations, leading to greater energy savings than reactive methods.
- Traffic_Factor is Key: Simply looking at packet arrival rates (PPS) is insufficient. Incorporating application processing requirements (CPP) into a single "Traffic_Factor" provides a more accurate measure of the actual processing demand.
- Queue Length as a Fine-Grained Indicator: While Traffic_Factor determines the number of active cores, the input queue length serves as a real-time indicator for fine-tuning the frequency of active cores using DVFS.
- DES Predictor Suitability: Double Exponential Smoothing is an effective and low-overhead predictor for network traffic, outperforming simpler methods and even complex ANN models in some aspects.
- DVFS Complements Core Management: Adjusting core frequencies (DVFS) provides substantial additional power savings (up to 39% in some cases) on top of managing the number of active cores.
Key Results
- The proposed PTM scheme achieves up to 40% more power savings compared to traditional power management schemes.
- PTM consistently results in the lowest average number of active cores across all benchmarks and traces.
- DVFS alone can improve power consumption by 15% to 39%.
- The DES predictor shows comparable or better accuracy than more complex predictors like ANN.
Limitations and Future Work
- Thresholds: The paper mentions that thresholds for queue length and other parameters can change dynamically. While this is a strength, finding optimal dynamic threshold adjustment strategies could be an area for further exploration.
- Global DVFS: The authors suggest that using global DVFS (all cores changing frequency in unison) instead of per-core DVFS could simplify the design while still offering significant power savings. This could be a direction for future work.
- Predictor Choice: While DES is effective, the paper notes that designers could use other predictors based on their specific requirements. Exploring the impact of different predictors on overall performance would be valuable.
Summary updated