Traffic imbalance among reverse proxies can make some of them very busy, leaving others underloaded. This can lead to an inefficient resource usage within the backbone.
In addition, when a request arrives at a proxy, it is possible that another proxy is much ``closer'' or has a less-crowded network connection to the client. Using this ``another'' proxy to service the request may yield better performance to clients. Current proxy cooperation cannot utilize this advantage.
While our proposed mechanism can potentially address all these issues, we concentrate on the issue of backbone traffic reduction in this paper. To understand how bandwidth is consumed for populating proxy caches, we analyzed the document access pattern on AT&T Worldnet IGRs using an Easy WWW trace. Figure 2 shows that the trace has two distinct peaks: more than 20% of the total number of bytes are only accessed once through any given IGR, while another 20-25% of bytes are referenced more than 50 times. All the documents accessed less than 5 times account for 40% of the documents. This interesting pattern leads to our observation that significant bandwidth savings could be obtained by reducing the bandwidth consumption for populating caches. Further, most of the documents fall into two distinct groups: one for which this bandwidth consumption is of no concern because it is amortized over many accesses, and the other which would benefit a great deal from addressing the issue.
The more bandwidth ISPs have, the more customers they can accommodate. Conversely, a crowded backbone may result in service degradation and client complaints. Therefore our goal in this paper is to reduce the bandwidth consumption used to populate reverse proxies with cached objects. Quantifying the benefits of our approach for the other two problems mentioned above remains future work.
