next up previous
Next: Changing the Time Interval Up: Experiments Previous: Second-level Proxies

Frequency of Top-10 Release

  In this section, we investigate how often a new Top-10 should be released. The Top-10 music charts have been released every week for several decades now, without any significant problems. It would be interesting to see if the same one-week time interval should apply to the calculation of the Top-10 of each Web server as well.

Intuitively we believe that the time interval for the Top-10 calculation should neither be too large, nor too small. For example, if a new Top-10 is released every several months, then it may be out of date, and all clients that prefetch it, may not use it. If, on the other hand, a new Top-10 is released every few minutes, then it will probably not be credible, and would imply a significant overhead for clients that would prefetch probably the same Top-10 every few minutes.

     figure251
Figure 13: Traffic Increase.
Figure 12: Hit Ratio.

To find what values of the time interval would be appropriate, we conducted a trace-driven simulation, where we vary the time interval from half a week up to a month, and plotted the performance results in figures 12 and 13 (TOP-10 is fixed at 500, with the exception of NASA where it as fixed at 100). We see that for most servers, Hit Ratio improves slowly with the time interval and then declines. gif Interestingly enough, we see that for all servers the best interval seems to be between one and two weeks. Thus, every one to two weeks, a new Top-10 should be released.

Figure 13 shows the traffic increase as a function of the time interval. We see that traffic increases slowly with the time interval, and sometimes it fluctuates around a value. For Parallab and Rochester, traffic increases sharply after two and three weeks, respectively. The reason is that after that time interval, lots of clients start to qualify for prefetching, and lots of prefetching operations start to clients that do not really need the prefetched data. Fortunately, for time intervals less or equal to 2 weeks, all servers' traffic increase is lower then 20%. Summarizing, the best balance between hit ratio and traffic increase seems to be achieved when a new Top-10 is released every couple of weeks.


next up previous
Next: Changing the Time Interval Up: Experiments Previous: Second-level Proxies

Evangelos Markatos
Fri Nov 1 16:38:26 EET 1996