Abstract
To improve the efficiency of resource utilization and scheduling of scientific data transfers on high-speed networks, the "Advanced Performance Modeling with combined passive and active monitoring" (APM) project investigates and models a general-purpose, reusable and expandable network performance estimation framework. The predictive estimation model and the framework will be helpful in optimizing the performance and utilization of networks as well as sharing resources with predictable performance for scientific collaborations, especially in data intensive applications. Our prediction model utilizes historical network performance information from various network activity logs as well as live streaming measurements from network peering devices. Historical network performance information is used without putting extra load on the resources by active measurement collection. Performance measurements collected by active probing is used judiciously for improving the accuracy of predictions.
Introduction
Modern scientific research relies on massive data transfers across high-speed networks like ESnet and Internet2, which now operate at speeds up to 100 Gbps. To schedule these transfers efficiently, engineers need to know two things: how much bandwidth is currently available and how much will be available in the near future.
Traditional tools rely on "active probing"—sending test packets across the wire to see how they perform. However, on 100 Gbps links, active probing is problematic. It can put an immense load on the network, potentially interfering with the very performance it's trying to measure, and it often fails to capture the nuances of ongoing background transfers. The APM project seeks to move toward "passive" monitoring, using the data routers already collect (like SNMP and NetFlow logs) to synthesize a predictive model of end-to-end performance.
Methodology
The paper details several novel algorithmic approaches to network modeling designed to handle the scale of modern backbones:
- Hybrid Forecasting (STL + ARIMA): Instead of using a standard ARIMA model (which is computationally expensive to train), the team used Seasonal-Trend decomposition by Loess (STL) to first strip out periodic patterns. They then applied ARIMA only to the remaining "noise." This decomposition allows for a more granular understanding of network behavior over time.
- Edge-to-Edge Throughput Inference: The researchers developed an algorithm that treats sudden spikes in SNMP link utilization as "events." By iteratively matching these spikes across neighboring routers in a topology, they can reconstruct the path and throughput of a specific data transfer without ever looking at the actual packet headers.
- Data Reduction (BO-LEMs): To handle the firehose of streaming network data, they implemented Bayesian Online-Locally Exchangeable Measures, a technique to reduce the number of samples needed for accurate analysis.
- Predictive Lasso: For NetFlow analysis, they utilized a "Predictive Lasso" method that calculates performance predictions in $O(np)$ time, avoiding the iterative complexity of traditional Expectation-Maximization (EM) algorithms.
Key Insights and Ideas
A central concept in this research is the "Traffic Analogy." The researchers view network bandwidth much like highway traffic. Just as a commuter knows a trip takes 90 minutes at 8:30 AM but only 40 minutes at 1:00 PM, network usage has "seasonal" patterns (daily/weekly cycles) that are highly predictable if decomposed correctly.
Furthermore, the study demonstrates that passive monitoring is sufficient. Router logs contain enough "echoes" of large transfers that the entire flow can be reconstructed by "connecting the dots" across the network topology. By separating the seasonal trend from the "remainder" noise, the forecasting model proved much more resilient to sudden, non-periodic spikes in bandwidth usage compared to traditional methods.
Datasets and Validation
The study utilized real-world data from major research backbones, including SNMP per-interface logs and NetFlow logs from ESnet and Internet2, as well as SNMP time series from Georgia Tech's border routers. To validate their inference algorithms, the team conducted controlled TCP data transfers between Georgia Tech and Lawrence Berkeley National Laboratory (LBNL) to compare predicted results against known ground truth transfer rates.
Results
The APM framework demonstrated significant improvements over traditional methodologies:
- Efficiency: The STL+ARIMA approach reduced computation time for model training by 78.1% compared to the traditional Box-Jenkins methodology.
- Accuracy: The edge-to-edge inference algorithm achieved over 95% recall for identifying transfers larger than 3 Mbps and longer than 2 minutes.
- Forecasting: Forecast errors were kept under 20% in most cases, and often hovered around 10%, even on high-bandwidth wide-area networks.
- Data Compression: The BO-LEMs algorithm reduced the required number of network measurement samples by 66% while maintaining accurate throughput prediction.
Conclusion and Future Work
While the results are promising, the authors note that real-time constraints remain a hurdle, particularly the frequency at which external SNMP servers can be queried. Additionally, the inference model's accuracy drops for very small flows (under 2-3 Mbps) as they become indistinguishable from background network noise. Future work involves expanding the framework's API so that middleware can automatically query these models to choose the best time and path for moving data.
Summary updated