Delayed Dissemination of Cache Content Information

 As mentioned earlier, we assumed so far that all proxies have instant, perfect knowledge of all cache locations of every requested document. We now consider how propagation delay of cache location information affects our results. To this end, we approximated the summary cache cooperation model [6]. In the simulations of this section, each proxy maintains a summary of other proxy's cache information. All proxies exchange this information at designated time points. In other words, at the designated time points, all reverse proxies get up-to-date information; this information then becomes stale and imprecise due to changes in each proxy cache content, until next exchanging time point. On a false hit i.e., when a proxy contacts a remote proxy that does not have the document, this proxy will get a negative acknowlegement from the remote one and then fetch the document from the content server itself.

Figure 10 shows the delay impacts on bandwidth consumption for the IGR scheme. There is a gap between the curves for the cases with propagation delays and the curve for the perfect knowledge case. We found that the gap mostly comes from incomplete location information, causing a forwarding proxy to pick sub-optimal remote proxies in some cases. False hits are pretty rare and only account for less than 0.06% of total requests on any proxy. Overall, our technique shows similar trends in bandwidth savings with delayed or with instant dissemination of cache location information. While the delays do affect the amount of bandwidth savings, the performance gap between perfect and imprecise knowledge is rather small: for bandwidth, the gap is up to 5% of total traffic, and for latency, the absolute difference is less than 1.5ms.

  

Figure 10: Impact of propagation delay of cache location information on bandwidth (IGR, 7-day-TTL).

  

Figure 11: Impact of propagation delay of cache location information on latency (IGR, 7-day-TTL).

Figure 11 shows the impact on latency in IGR. It shows that adding delays reduces latency overhead, compared to perfect knowledge. The reason for lower latency overhead is as follows: due to the propagation delay, a proxy receiving a request sometimes does not know that the requested object is cached on another proxy. The first proxy would then fetch the document from the content server rather than forward the request to the remote proxy. Thus, request forwarding is applied less frequently; therefore the latency overhead of request forwarding is reduced, and the reduction is accompanied by the corresponding reduction in bandwidth savings, as seen on Figure 10.

As for Server-Forwardable, where content server joins this forwarding mechanism, from Figure 12 and Figure 13, we see practically no difference between different cases. So, our mechanism is robust against information propagation delays when the server accepts forwarded requests.

  

Figure 12: Impact of propagation delay of cache location information on bandwidth (Server-Forwardable, 7-day-TTL).

  

Figure 13: Impact of propagation delay of cache location information on latency (Server-Forwardable, 7-day-TTL).