Performance Analysis of a Dynamic Parallel Downloading Scheme from Mirror Sites Throughout the Internet

Allen Miu, Eugene Shih
6.892 Term Paper, December 1999.

In this paper, we describe paraloading, a novel approach to retrieving files from the Internet by using parallel connections opened to multiple mirror sites.  This is in contrast to the conventional download method where the client retrieves data from a single source. The advantages of paraloading over single-connection downloading include improved performance due to the aggregation of the bandwidths of the parallel connections, the improved resilience to occurrences of congestion or failures on any one path, and the ability to perform load balancing among the parallel connections. In order for paraloading to work, servers must be mirrored throughout the Internet.  However, as mirror sites become more widespread and as end users upgrade to higher connection speeds, we believe that paraloading--if implemented properly--will offer significant performance gain over the traditional single source file access.

Paraloading is a subject that has not been fully studied by the research community.  This paper examines the performance and the design of a paraloading scheme proposed by Rodriguezet al.  We have developed a paraloader application in Java that uses HTTP 1.1 for its range-request and persistent connection features.  We have conducted a series of experiments using our paraloader at MIT and at UC Berkeley, and we have found that the performance gains of paraloading are not as good as those claimed by Rodriguezet al.  This suggests that paraloading may not fair well in different network environments.  Nevertheless, we believe that there are a number of enhancements that can be made to the paraloader to improve its performance in different network environments. We will outline some of these enhancement techniques and discuss some open research issues on paraloading.

Please be advised that the paper is still in its preliminary stages.

[Gzipped PostScript (78KB)][PDF (160KB)][MS PowerPoint Presentation (630KB)]