We found that Web++ improves the response time during the peak hours on the average by 47.8%, at least by 27.8% and at most by 59.1% when compared to the single any single server system. At the same time, it degrades the response time relative to the optimal system on average by 2.2%, at least by 0.2% and at most by 7.1%. Not surprisingly, we found that the performance benefits of Web++ are weaker during the evening hours. In particular, we found that Web++ improves the response time on the average by 25.5%, at most by 58.9% and in the worst case it may degrade performance by 1.4%. We also found that Web++ degrades the response time with respect to the optimal system on average by 25.5%, at least by 7.8% and at most by 31%. Throughout all the experiments we found that Web++ did not send more than 6 extra requests to refresh its latency table (compared with three times as many requests sent by the optimal system!).

The experimental results can be found in Figures 11
- 14. Figures 11 and
13 show the median of the response time during office
and evening hours. Figures 12 and
14 show the relative median of the response time with
respect to the median response time of the optimal system. Compared
with the experimental results reported in [35] (an average
69% improvement in HTTP request *latency*), the results
reported here (an average 37% improvement in HTTP request *response time*) indicate a weaker improvement in performance. We
believe that the difference is due to a smaller number of replicas for
each resource in the experiments reported here (3 compared with 5 and 50 in
[35]). The bigger the number of replicas the higher the
probability that a client finds a replica ``close'' to it.