Project Title: Bandwidth Estimation: Measurement Methodologies and Applications
Project Type: SciDAC
PI:Constantinos Dovrolis, University of Delaware
1 Executive summary
The major goal of our research is to develop bandwidth estmation (bw-est) techniques and tools that will allow scientists to use effectively high-performance network infrastructures and demanding distributed applications.
The objective that we had set for the first six months of this project was to develop an available bandwidth measurement methodology, and to implement that methodology in a software tool. That objective has been met. Specifically, we have recently developed an original measurement methodology, called Self-Loading Periodic Streams (SLoPS). To the extent of our knowledge, SLoPS is the only existing technique that can measure end-to-end available bandwidth without saturating the network (i.e., non-intrusively). We have implemented SLoPS in a tool called pathload, and we have tested and verified pathload in several paths within the US and Europe.
Over the last six months, we have been also improving our (previously implemented) capacity estimation tool, called pathrate. Pathrate is currently being used by other DOE researchers. After the SciDAC meeting in 1/2002, we have started collaborating with several researchers in extending, improving, and distributing our measurement software.
2 Current accomplishments
The major progress over the last six months was the development of the Self-Loading Periodic Streams (SLoPS) measurement methodology. SLoPS measures end-to-end available bandwidth in a fast, accurate, and most importantly, non-intrusive manner, i.e., without saturating the path. We have implemented SLoPS in a measurement tool called pathload. We have tested and verified the operation and accuracy of the tool in several paths that cross the United States and Europe.
We describe SLoPS and pathload in the following two technical reports:
- "Pathload: a measurement tool for end-to-end available bandwidth", by Manish Jain and Constantinos Dovrolis. To appear at the Passive and Active Measurements (PAM) workshop, Fort Collins, Colorado, March 2002. Available at: http://www.cc.gatech.edu/fac/Constantinos.Dovrolis/Papers/pam02.ps
- "end-to-end available bandwidth: measurement methodology, dynamics, and relation with TCP throughput", by Manish Jain and constantinos Dovrolis. Submitted for publication to the 2002 ACM SIGCOMMConference. Available at: http://www.cc.gatech.edu/fac/Constantinos.Dovrolis/Papers/sigcomm02.ps.gz
We are currently working on a first release of pathload (source code). Our objective is to do so in Spring 2002.
The development of SLoPS has been a significant progress in the bw-est state-of-the-art. This is because, prior to SLoPS, it was not known how to measure available bandwidth in an end-to-end manner. Tools such as pathchar, pchar, and pathrate, attempt to measure capacity, and not available bandwidth. Other tools such as pipechar and cprobe attempt to measure available bandwidth using packet trains. In a recent publication, however, we have shown that such techniques do not measure available bandwidth . Finally, tools such as Iperf or BBFTP, attempt to measure the available bandwidth using bulk TCP connections, saturating the path and affecting its cross traffic. To the extent of our knowledge, SLoPS is the only existing methodology that can measure end-to-end avaialble bandwidth in a non-intrusive manner.
In parallel with our work on available bandwidth estimation, we have been working on capacity estimation. Prior to the start of this project, the PI had written another tool, called pathrate, that measures the end-to-end capacity of a path. Over the last six months, we have been working on improving pathrate. The results of this work can be found at the following report:
- "Packet dispersion techniques and capacity estimation", by Constantinos Dovrolis, Parmesh Ramanathan, and David Moore. Submitted for publication to the IEEE/ACM Transactions on Networking. Available at: http://www.cc.gatech.edu/fac/Constantinos.Dovrolis/publications.html
The source code for pathrate is available in public-domain at: http://www.pathrate.org.
3 Future accomplishments
Over the next six months, we will focus our work on the following tasks:
- release of pathload: this includes porting the code to all mainstream versions of Unix, increasing its robustness to user or other errors, providing a more friendly user-interface, and supporting the tool. We plan to release the code in three stages. The first release will be to a dozen of colleagues that work on bw-est. Their feedback will help us to significantly improve the tool. In a second release we will announce the tool at the IETF IPPM working group and other related forums and conferences. The third release, probably before the end of 2002, will include announcements at large mailing lists, and an article at the USENIX login.
- We are working on an experimental study of the statistical characteristics of available bandwidth. Specifically, we attempt to measure the predictability and variability of available bandwidth in different timescales and operating conditions (network load, degree of statistical multiplexing, distribution of round-trip times, etc.). these statistical characteristics are of major importance for applications and transport protocols. The reason is that depending on how predictable the available bandwidth is, applications can configure their measurement rate, and the confidence interval of the resulting esimates. We plan to submit a research paper describing this work at the SIGCOMM International Measurement Workshop (IMW) in June 2002.
- We are currently working on a research paper that shows the limitations of the Variable Packet Size (VPS) bw-est methodology. VPS is used by pathchar, pchar, and clink. We have recently discovered that VPS-based tools consistently underestimate the per-hop link capacities, when the network paths include layer-2 switches. Given the commodity of such layer-2 technologies, the VPS methodology cannot be expected to provide accurate results. We plan to submit this paper to the ACM Computer Communications Review (CCR).
- We are also writing a survey paper on bw-est methodologies and tools. Given the significant research interest in this area over the last couple of years, it is important to educate the networking community about the major recent results, the characteristics and limitations of each tool, and the remaining research challenges. We plan to submit this paper to the IEEE Network magazine.
4 Research interactions
Over the last six months, we have collaborated with several other DOE researchers, including Les Cottrell (SLAC), Tom Dunigan (ORNL), Jin Guojun (LBNL), and of course with our co-PI K. Claffy (CAIDA).
Specifically, we work with Les Cottrell on improving pathrate, testing it over a large number of national and international paths, and comparing its results with other similar tools. We plan to extend this collaboration with pathload experiments, after we release that tool.
Our collaboration with Tom Dunigan has also focused on pathrate. Tom has tested pathrate in several high-performance paths, limited by OC3 aned OC12 links. These experiments are crucial for the verification and tuning of the tool, given that we do not have access to such high-performance links. Tom has also tested pathrate in paths that include traffic shaping devices.
We recently started working with Jin Guojun on a first draft of a measurement middleware architecture, based on bw-est.
After the SciDAC meeting in January 2002, we have set the groundwork to collaborate with Wu-chen Feng (LANL), Rudolf Diedi (Rice), Miron Livny (UWisc), Karsten Schwan (GATech), Deb Agrawal (LGNL), and others.
This projects is joint work with the graduate students Manish Jain and ravi S. Prasad, both funded by this SciDAC project at the University of Delaware. We are also collaborating with our CAIDA colleagues: K. Claffy (co-PI), Margaret Murray, and David Moore.
 C. Dovrolis, P. Ramanathan, and D. Moore, "What do Packet Dispersion Techniques Measure?," in Proceedings of IEEE INFOCOM, Apr. 2001.