Architectures emulated on test bed

In our experiments, we were unable to stress web server with heavy load due to limited available RAM (32 MB) and client machines were not able to handle heavy data rate or run large number of webstone processes. Due to packet delay software (which was run as kernel module), when data rate was high, buffering large amount of data for delay period consumed more RAM and generated very high interrupt rate and Linux kernel did not handle the situation gracefully. Even kernel compiled with option "CPU is too slow to handle full bandwidth" did not make systems stable when the data rate was high. Due to these limitations, we could not create the situations when queuing or processing delays at server dominate network delays.

We have emulated four policies for cluster selection at DNS in our test bed and we discuss the results obtained for those policies below: