Roger P. Karrer
Senior Research Scientist
Address:
Ernst-Reuter-Platz 7
D-10587 Berlin
Germany
Phone:+49 30 8353 58459
Fax: +49 30 8353 58409
Email: roger [dott] karrer [att] telekom [dot] de
Parallel download protocols for the Internet
Introduction
Parallel download protocols offer a decrease in the download time for a single user compared to single-flow downloads, resilience to slow performance along a path or even path failures, and a system-wide load balancing by distributing the download over multiple servers. Therefore, parallel download protocols are are increasingly used to overcome limitations of the Internet as a large distributed system. While many parallel download protocols have been deployed and are extensively used, my objective is to design, implement and experimentally evaluate high-speed parallel download protocols from multiple replicas to one client and assess performance, fairness and real-time streaming properties:- TCP-PARIS: the world's fastest parallel download protocol
- TCP-ROME: a protocol framework that allows a trade off agressivity and fairness
- Future work
TCP-PARIS: the world's fastest parallel download protocol.
Parallel Download Protocol for Replicas is the world's fastest parallel download protocol, subject to TCP friendliness. Our paper describes that TCP-PARIS is a transport-layer parallel download protocol. We argue that a transport-layer approach has at least two significant advantages. First, content in the Internet is increasingly distributed over multiple servers: Content Distribution Networks, Mirror Servers, and, most recently peer-to-peer systems. Currently, all these applications use different parallel download protocols. TCP-PARIS alleviates the overhead of deploying different parallel download protocols with a single solution. TCP-PARIS thereby requires only that a TCP-PARIS socket instead of a TCP socket be instantiated - the remaining parallel download is hidden from the application. Second, TCP-PARIS uses transport layer only to determine which and how much data be downloaded from each server. In particular the congestion window information allows TCP-PARIS to achieve a near-to-optimal performance independent of file size, bandwidth heterogeneity and bandwidth changes during the download. Our paper shows with simulations of typical peer-to-peer scenarios that TCP-PARIS achieves up to 52% download time reductions with 8 parallel connections compared to a download with a single connection, and up to 52% performance gains compared to related protocols, such as BitTorrent and DigitalFountain.Resources
- R. Karrer and E. Knightly, TCP-PARIS: a Parallel Download Protocol for Replicas. In Proceedings of IEEE International Workshop on Web Content Caching and Distribution (WCW'05), Sophia Antipolis, France, September 2005 pdf
- Slides from the talk at WCW'05.
Ns-2 code
Available soonImplementation for the Linux kernel
Coming soonTCP-ROME: mitigating fairness and performance in parallel download.
Parallel download protocols can reduce the download time of an application by up to a factor of 1/n when using n parallel connections in a heterogeneous environment. However, the performance gain for a parallel download can unfairly decrease that of other applications: single connection applications running on the same host, applications that share the same network bottleneck or applications that connect to only one server. TCP-ROME is a protocol framework that allows an application to adjust the aggressiveness of a parallel download protocol. TCP-ROME can be implemented as an extension of TCP-PARIS. This work is currently under submission. More information will be available soon.Ns-2 code
Available soonImplementation for the Linux kernel
Coming soonFuture work
We are currently extending our work on TCP-PARIS and TCP-ROME in the following direction.- Application integration. A key question for
parallel downloads is how many connections should be used for the parallel
download. While simulations indicate that the biggest performance gain
when moving from 1 to 2 parallel connections, we want to deploy TCP-PARIS
and TCP-ROME and perform a measurement study in the Internet.
- We are exploiting the impact of parallel download protocols on the Linux kernel. In contrast to application-layer protocols, TCP-PARIS and TCP-ROME are limited by the amount of kernel buffer. We are investigating the constraints and develop solutions for them.