Bandwidth Estimation: Measurement Methodologies and Applications

Kimberly Claffy, CAIDA/SDSC/UCSD
Constantinos Dovrolis, Georgia Institute of Technology

Summary

Internet resources are often used inefficiently due to the inability of high bandwidth applications to accurately ascertain bandwidth parameters in real-time. This SciDAC project strives to improve existing bandwidth estimation techniques and tools, and to test and integrate the most successful of these technologies into DOE and other network infrastructures.

More accurate and efficient bandwidth estimation tools and techniques would enable high performance network engineering and management functionality that have remained elusive for years, specifically the capabilities to:

•  Automatically size TCP socket buffers to achieve better throughput.
•  Configure overlay routes for overlay networks.
•  Select the highest throughput server for content distribution.
•  Adjust the encoding rate on streaming applications.
•  Monitor path load to verify SLA and QoS targets.
•  Check for sufficient bandwidth as part of application admission control.
•  Construct application level topologies for peer-to-peer networks.
•  Select an egress ISP for interdomain traffic engineering.

Our specific progress towards meeting these needs includes:

Survey of Current Tools: “Bandwidth Estimation: Metrics, Measurement Techniques, and Tools”, published in IEEE Network magazine (Nov/Dec 2003), presents a taxonomy of currently available tools, comparing their main differences and similarities.

Refinement of Tool Testing methodology: We improved our lab environment and configuration and devised better methods for generating realistic yet reproducible simulated cross-traffic. We also automated test data collection and improved our capabilities for independently measuring and graphing cross-traffic and tool traffic using a NeTraMet passive monitor.

Figure 1. Tool Traffic, Cross-traffic, and Passive Traffic Monitoring in the Bandwidth Estimation Testbed. Tool traffic is shown between end hosts in the upper part of this figure. Several options for generating cross-traffic loads are illustrated in the lower part of the figure. All three routers connect to the Spirent SmartBits 6000 at the center bottom. A PC running the Spirent SmartFlow application controls the direction, makeup, and load level of multiple logical traffic flows on any or all physical links. Finally, a NeTraMet passive meter (top right) runs a ruleset to track packet sizes and inter-arrival times from all flows.

Analysis of E2E bandwidth estimation tools on both 100Mbps and GigEther links: Several factors affect tool accuracy including: the presence of layer-2 store and forward devices; differences in the size of internal router queues; and high cross-traffic loads. All of these conditions are likely to occur in the Internet in the wild, complicating the task of end-to-end bandwidth estimation.

Testbed sharing with other DOE bandwidth estimation researchers: We opened up our bandwidth estimation test lab for use by DOE collaborators.

Application of Internet spectroscopy techniques to link capacity characterization: We began to investigate a new technique for revealing characteristics of layer-2 technologies without requiring additional traffic probes. Internet spectroscopy is based on an algorithm where a radon transform of inter-packet delay distributions is coupled with entropy minimization.

Pathrate and Pathload in gigabit paths: New versions of our bandwidth estimation tools, Pathrate and Pathload, were released in January 2004. Pathrate v2.3.3 measures end-to-end capacity (a.k.a., bottleneck bandwidth), while Pathload v1.1.1 measures end-to-end available bandwidth.

Socket buffer sizing for maximum TCP throughput: Bandwidth estimation can significantly improve the throughput of large TCP transfers, such as the transfers of large scientific data sets. We have developed an application-layer technique, called SOcket Buffer Auto-Sizing (SOBAS) , that helps TCP to achieve its maximum feasible bandwidth in a network path. SOBAS does not require changes in TCP. The key idea behind SOBAS is to limit the socket buffer size, and thus the maximum TCP send-window, to the point that the transfer saturates the network path without causing packet losses.

Figure 2. TCP throughput is much higher when using SOBAS (upper graph). SOBAS maintains the optimal socket buffer size at maximum lossless receive throughput.

Plans: This project ends in August 2004. We hope to leverage the investment DOE has already made in improving the integrity of the component technologies in a follow-on effort. To this end, we are developing a proposal in collaboration with ORNL and the Pittsburgh Supercomputer Center (PSC) for Pythia, an automatic performance monitoring and problem diagnosis system for ultra high-speed networks.

For further information on this subject contact:
Dr. Margaret Murray, Program Manager
Cooperative Association for Internet Data Analysis (CAIDA) at the San Diego Supercomputer Center
University of California, San Diego
Phone: (858) 534-8928 margaret@caida.org
Project URLs:
http://www.caida.org/projects/bwest/
http://www.pathrate.org/

back to project page