Abstract
Content Delivery Networks (CDNs) face significant challenges during traffic spikes where demand exceeds server capacity. While Peer-to-Peer (P2P) offloading is a common solution, excessive offloading can lead to underutilized infrastructure and reduced operator profit. This paper presents the Dynamic Moderate Peer-assisted Method (DMPM), which utilizes ARIMA-based time-series forecasting to predict traffic and offload only the excess demand to a P2P overlay. Evaluation using real-world traces from ChinaCache demonstrates that DMPM maintains high Quality of Service (QoS) while maximizing CDN server utilization.
Introduction
CDNs handle massive amounts of web traffic, but their server capacity is finite. When flash crowds occur, demand can exceed capacity, leading to dropped requests and poor Quality of Service (QoS). To solve this, hybrid CDN-P2P systems were developed to offload excess traffic to end-users (peers).
However, previous research focused primarily on maximizing P2P offloading. From a CDN operator's perspective, this is economically flawed. Operators monetize the traffic served by their own infrastructure; if P2P handles too much traffic, expensive CDN servers sit idle, and resource utilization plummets. The authors identify a critical gap: how can a CDN operator know exactly when and how much traffic to offload to a P2P network so that CDN servers remain at near 100% utilization without sacrificing user QoS?
System Architecture
The proposed system utilizes a hybrid architecture where individual CDN nodes, organized in star topologies, can form a P2P overlay to share the load during peak demand periods.
The Dynamic Moderate Peer-assisted Method (DMPM)
The authors propose the Dynamic Moderate Peer-assisted Method (DMPM), a distributed algorithm running on individual CDN nodes without centralized control. It consists of two main components:
- Online Load Prediction (ARIMA): Instead of reacting to traffic spikes after they happen, DMPM predicts aggregated resource demand one time-step into the future. It uses an Autoregressive Integrated Moving Average (ARIMA) model. The algorithm automatically tests the time-series data for stationarity, applies differencing if necessary, and uses the Bayesian Information Criterion (BIC) to select the optimal parameters dynamically.
- Popularity-Aware P2P Policy: The system sets a Warning Value (e.g., 85% of maximum capacity). If the ARIMA model predicts that the next time slot's demand will exceed 100% capacity, DMPM calculates the exact overflow amount.
The Moderate Offloading Strategy
The core philosophy of DMPM is "moderate" offloading. Rather than offloading as much as possible, the system identifies the specific time window where demand exceeds the rated capacity of the CDN servers.
As shown in the figure above, when demand (the black curve) exceeds the capacity (the red line) between points $t_1$ and $t_2$, the system identifies the most popular content currently being requested. It then uses the View Upload Decoupling (VUD) method to form a P2P overlay specifically for that content, offloading only enough requests to bring the CDN server load back down to its maximum rated capacity.
Evaluation and Results
The authors evaluated their method using real-world traffic traces from ChinaCache, a major CDN provider in China. The dataset consists of 7 continuous days of aggregated request logs from September 2010, collected from approximately 95 geo-distributed CDN server nodes.
Performance was measured using Prediction Error (PE), CDN Resource Utilization (CRU), and Request Response Rate (RRR):
- High Prediction Accuracy: For 6 out of the 7 days tested, the ARIMA model achieved a Prediction Error of less than 0.1 (10%).
- Maximized Utilization: During peak times when demand exceeded capacity, DMPM successfully kept the CDN Resource Utilization (CRU) at nearly 100%.
- Maintained QoS: The Request Response Rate (RRR) remained at 1.0 (100% of requests served) during normal and moderately high traffic, proving the P2P overlay successfully absorbed the overflow.
- The "Thursday" Anomaly: On one specific day featuring a massive, sudden flash crowd (demand spiked to nearly 2x capacity instantly), the ARIMA model failed (PE > 0.2), which caused a temporary drop in QoS.
Conclusion and Future Work
By treating P2P strictly as an overflow valve rather than a primary delivery mechanism, CDNs can align technical scalability with economic incentives. While DMPM performs well under standard daily traffic patterns, its primary limitation is vulnerability to extreme, sudden flash crowds that ARIMA cannot accurately forecast. Future work could explore non-linear forecasting methods or machine learning approaches to better anticipate and react to these extreme, bursty demand spikes.
Summary updated