SPAWN (Scalable Proxy Architecture for web access over Slow Wireless Network) is a dual-proxy driven architecture designed to overcome delays associated with Web access over slow wireless networks by combining a variety of known and new techniques, including compression and differential compression of HTML pages, transcoding of images, caching of old and similar content, and optimization of the transmission protocol (HTTP) to avoid extra roundtrips. SPAWN’s proxy server is based on the single-process event driven architecture for maximum concurrency. The event-driven approach implements the processing of each operation as a finite state machine, where various events either system or user generated (e.g. readiness or completion of a network I/O) trigger transitions between states, allowing an application specific centralized control over all the events. SPAWN makes use of this architecture to implement intelligent scheduling decisions for efficient resource allocation (e.g. CPU).

We provide a systematic performance study of exactly what happens when such a proxy server runs under overload and identify resources that experience overload first among others. The following conclusions are drawn from different experiments:

1. We have shown that the CPU-intensive image transcoding operations are the primary cause for overloading and a transcoding policy that considers system load as one of the criteria while making transcoding decisions performs better than a policy that does not.

2. Contrary to popular thoughts that transcoding should be performed if it reduces latency, results from the experiments suggest that it is not always true.

3. Both RR (Round-Robin) and EDF (Earliest-Deadline-First) scheduling policies perform similar except SPT (Shortest-Processing-Time), which suffers from starvation problem and hence perform severely bad, in the case when transcoding subsystem does not take client into consideration while making transcoding decisions.

4. We have shown the impressive performance improvements under "scalable selective transcoding" as compared to "non scalable selective transcoding", i.e., the benefits of considering client load while making transcoding decisions. Also, an EDF scheduling policy provides further performance improvements over the RR scheduling policy.

This project was supported by a grant from Intel.

Relevant Publications

1. Anubhav Savant. SPAWN: a Scalable Proxy Architecture for web access over slow Wireless Networks. MS Thesis, Department of Computer Science, Polytechnic University, 2003.

2. Anubhav Savant, Nasir Memon and Torsten Suel. On the Scalability of an Image Transcoding Proxy Server. In IEEE International Conference on Image Processing, September 2003, Barcelona, Spain. PDF 122KB