 |
|
 |
||||||
|
||||||
 |
Home
| Mission
|
about SciDAC
|
Contact Us |
|||||
 |
 |
 |
||||
|
Alumni ProjectBandwidth Estimation: Measurement Methodologies and Applications
Dennis Gannon, PI. Department of
Computer Science, Indiana University SummaryInternet resources are often used inefficiently due to the inability to cooperatively share and access existing bandwidth. High bandwidth applications such as grid interfaces, high-speed file transfers, and streaming media are particularly hindered by the inability to accurately ascertain bandwidth parameters in real-time. This 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. Bandwidth estimation tools are especially important to high throughput data-intensive distributed applications. Scientific visualization, interactions within distributed collaboratories, and remote sensor analysis are examples of applications that typically push the edge of network performance. The scientific community can substantially benefit from better bandwidth estimation techniques to optimize their application performance. This project seeks to:
As part of this SciDAC project, we have developed software tools for measuring the capacity and available bandwidth of Internet paths. Capacity is the maximum throughput that a network path can offer to an application (see Figure 1). Our capacity estimation tool is called pathrate. Available bandwidth, on the other hand, is the part of a path's capacity that is not utilized by competing cross-traffic. Our available bandwidth tool is called pathload. Both tools can monitor the bandwidth of a path without saturating it with measurement traffic. These tools are being used as sensors in SLAC’s IEPM-BW infrastructure. The algorithms of pathrate and pathload are based on novel measurement techniques, published (after peer review) in the prestigious IEEE INFOCOM 2001 and ACM SIGCOMM 2002 conferences. Figure 1. pathrate measures capacity, the maximum possible end-to-end throughput. pathload measures available bandwidth, or the maximum throughput given the current cross-traffic load. Both tools are available for download at http://www.pathrate.org . We compared different bwest tools using simulated cross-traffic in several 100Mbps and GigEther configurations. Results reveal that many existing tools do not work well in common network topologies. The presence of layer-2 store-and-forward devices causes consistent errors in per-hop measurement tools such as pathchar and pchar. Also, not all routers along an end-to-end path are equal: internal switches, different buffer configurations, and the relegation of classes of messages typically employed by tools to a router’s “slow path” all add error into the measurements. Furthermore, tool accuracy tends to deteriorate on heavily loaded or high bandwidth paths where traffic characteristics differ from tool assumptions or network interface interrupts coalesce. Attempts to increase accuracy often require tools to saturate the path with measurement probes, a method that is inefficient and does not scale. We are also investigating a new bandwidth estimation technique - Internet spectroscopy - that works by observing cell or slot-based traffic on broadband links and does not require additional probes. Internet spectroscopy is based on an algorithm where a radon transform of inter-packet delay distributions is coupled with entropy minimization. Layer 2 technologies such as ATM, rate-limited ATM, DSL, PPP, Ethernet and cable modems generate subtle yet characteristic features in IP traffic from which we can identify provisioned bandwidth. We plan to move network spectroscopy forward from a computational art into a viable tool for managing distributed applications. In order to enable automated analysis, we plan to build a library of delay spectra corresponding to known devices and link types. Such a library will enable us to distinguish variations that may correlate to specific markets or providers. We need to assess the accuracy of spectroscopy-based inference, checking whether measurement artifacts can distort the data interpretation. Finally, in an effort to improve communication and encourage collaboration with other DOE and networking community research, we plan to hold additional bandwidth estimation workshops for interested researchers and application developers. Our first workshop last June brought together researchers from LBL, SLAC, UC Santa Barbara, the University of Manchester, SDSC ‘s Enterprise Network Systems group and CAIDA. The summary report from that workshop is available at: http://www.caida.org/projects/bwest/presentations/mtgjun02/summary.xml . 
Figure 2. Testing against increasing cross-traffic loads reveals that pipechar (LBL) measures capacity while pathload (GATech) and netest-2 (LBL) reflect available bandwidth. For further information on this subject contact:
|
||||
Home | ASCR | Contact Us | DOE disclaimer |
|
|
