INCITE: Edge-based Traffic Processing and Inference for High-Performance Networks

Richard G. Baraniuk, Wu-chun Feng, Les Cottrell, Edward W. Knightly, Robert Nowak, Rolf Riedi2005 · DOI

TL;DR

This paper details the INCITE project's development of a suite of tools for understanding and managing high-performance networks. By focusing on edge-based measurements and advanced statistical analysis, INCITE provides novel methods for traffic characterization, path inference, and network monitoring, enabling more network-aware applications and improved troubleshooting.

What Problem Does This Paper Address?

Modern high-speed computer networks, distributed systems, and computational grids are incredibly complex. Optimizing their performance requires end-systems to understand internal network conditions. However, without specialized support at every router, this information is hard to obtain. The INCITE project tackles this by developing tools that infer dynamic network characteristics from measurements taken at the network's edge (hosts or edge routers). This addresses the need for adaptable, real-time analysis and control mechanisms that can predict network and application behavior based on limited data, moving beyond ad-hoc or overly restrictive existing approaches.

Methods

The INCITE project is structured around three key research thrusts:

  • Multiscale Traffic Analysis and Modeling: This thrust focuses on developing sophisticated models to capture the complex, bursty nature of high-speed network traffic. They leverage wavelets and multifractal theory to create statistically robust and analytically tractable models. Novel models like the multifractal wavelet model (MWM) and the alpha/beta model are introduced, which decompose traffic into components to better understand burstiness and the impact of protocols like TCP.
  • Inference and Control Algorithms for Network Paths, Links, and Routers: This area involves developing tools to understand network behavior from end-to-end measurements. Key techniques include:
  • End-to-end path modeling and inference: Using multifractal models to understand path dynamics.
  • Network tomography: Estimating performance (like loss and congestion) on internal network links using only external measurements. This includes Network Delay Tomography and methods for Merging Inferred Network Topologies.
  • New protocols for large file transfers: Developing protocols like TCP-LP and HSTCP-LP that efficiently utilize excess bandwidth without disrupting existing traffic, and TCP-Africa which adapts its aggressiveness based on network utilization.
  • Active network measurement and testing: Tools like pathChirp (using exponentially spaced chirps for bandwidth estimation) and STAB (Spatio-Temporal Available Bandwidth estimator) are developed.
  • Passive application-level traffic measurement: Tools like MAGNET (Monitoring Apparatus for General kerNel-Event Tracing) and TICKET (Traffic Information-Collecting Kernel with Exact Timing) are created to capture network traffic at the application level, before it's modulated by the protocol stack.
  • Data Collection Tools: This thrust underpins the others by ensuring data is collected and made available.

Datasets

Not explicitly detailed as distinct datasets for training or evaluation. The paper describes using real-world Internet traffic traces and measurements from various network testbeds (e.g., CAIDA Gigabit testbed, ESnet, Abilene/CENIC). The tools developed are intended for use on live network traffic.

Key Insights and Ideas

  • Edge-based inference is crucial: For complex networks, inferring internal states from edge measurements is a viable and necessary approach when direct internal access is impossible.
  • Traffic is highly bursty and multiscale: Simple models are insufficient; advanced statistical techniques like multifractals are needed to accurately capture network traffic dynamics.
  • A small fraction of connections cause most bursts: The alpha/beta model reveals that a small percentage of connections are responsible for the majority of traffic burstiness, which has significant implications for modeling and management.
  • Network topology changes significantly impact performance: The correlation between route changes and bandwidth fluctuations is substantial, highlighting the need for tools that can detect and diagnose these events.
  • Passive monitoring at the kernel level is powerful: Tools like MAGNET can capture traffic before protocol stack modulation, offering a clearer view of application behavior and the impact of lower network layers.
  • New transport protocols can improve efficiency: Protocols like TCP-LP and HSTCP-LP demonstrate that it's possible to utilize network resources more effectively for background tasks without negatively impacting interactive traffic.

Key Results

  • Development of novel traffic models: The MWM and alpha/beta models provide more realistic and analytically tractable representations of network traffic.
  • Creation of a suite of measurement and analysis tools: This includes pathChirp, STAB, AbwE/abing, NetRadar, MAGNET, TICKET, and TCP-LP/HSTCP-LP, many of which are released as open-source.
  • Demonstrated effectiveness in identifying network issues: Tools were used to identify performance degradation due to route changes, as illustrated by the SLAC to Caltech path example.
  • Significant performance improvements in monitoring tools: MAGNET showed a 50%-100% improvement over previous versions and was more unobtrusive than tcpdump.
  • New protocols show promise: HSTCP-LP was shown to utilize excess bandwidth effectively with minimal impact on other traffic.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that some approaches are still in the research and testing phase (e.g., network tomography tools).
  • While tools are designed for high-speed networks, achieving accurate measurements on very high-capacity links (OC-12 and above) can still present challenges due to issues like packet reordering and framing.
  • The cost-effectiveness of TICKET is highlighted, but it requires two PCs for software and hardware components.
  • Future Work:
    • Integrating INCITE tools with other SciDAC projects and infrastructure (e.g., Globus MDS, ESnet NIMI).
  • Further development and validation of network tomography tools.
  • Exploring the use of traffic modeling/synthesis tools for accurate performance predictions for computational Grids.
  • Making more detailed comparisons between different packet dispersion tools on high-capacity links.
  • Developing IPv6 versions of tools.
  • Continued research into the effects of new technology devices on network performance.

Summary updated

Abstract

The INCITE (InterNet Control and Inference Tools at the Edge) Project developed on-line tools to characterize and map host and network performance as a function of space, time, application, protocol, and service. In addition to their utility for trouble-shooting problems, these tools will enable a new breed of applications and operating systems that are network aware and resource aware. Launching from the foundation provided our recent leading-edge research on network measurement, multifractal signal analysis, multiscale random fields, and quality of service, our effort consisted of three closely integrated research thrusts that directly attack several key networking challenges of DOE's SciDAC program. These are: Thrust 1, Multiscale traffic analysis and modeling techniques; Thrust 2, Inference and control algorithms for network paths, links, and routers, and Thrust 3, Data collection tools.
BibTeX
@misc{baraniuk2005incite,
  author = {Richard G. Baraniuk and Wu-chun Feng and Les Cottrell and Edward W. Knightly and Robert Nowak and Rolf Riedi},
  title = {{INCITE: Edge-based Traffic Processing and Inference for High-Performance Networks}},
  year = {2005},
  doi = {10.2172/876141},
}
Figure