There are many limitations in our study. First, we assume fixed user request arrival times in our simulation. In practice, as requests are serviced quicker, the requests also arrive faster. Thus, the latency reduction discussed only serves as an indication of how much pre-pushing schemes can speed up user requests. Second, our calculation of client latency is merely an estimate based on the timestamps recorded in the traces and the modem bandwidth. It does not include protocol overhead and is far from accurate. However, we feel that the estimate is good enough to give us an indication of how well the pre-pushing scheme might perform.
Third, our simulator does not model the proxies accurately. In practice, the latency that a proxy incurs to fetch a document from its cache depends on many factors, including whether the document is cached in main memory, the load on the proxy, and the number of disk arms in a proxy. Our simulator takes the simplifying assumption that all cached documents are on disk, there is only one disk arm, and the proxy always takes 10ms to fetch a document. Clearly, only implementation experiments can demonstrate the real performance of pre-pushing schemes, and we are currently working on an implementation.
Finally, since the traces we use do not mark whether a response from the Web server carries cookies or not, our simulation considers all responses cacheable at the proxies. This inflates the proxy cache hit ratio. In addition, our simulation does not consider document changes, and treat modified documents as cache hits. This also has the effect of inflating proxy cache hit ratio. On the one hand, high proxy cache hit ratio may make our latency reduction appear higher. On the other hand, there are other techniques such as delta compression that can reduce client latency for modified documents, which we did not simulate. Thus, we believe that the results reported here still serve as a good indication of achievable latency reduction from the pre-push technique.