Selecting the Client Workloads

Next: Improvements from Route Control Up: Experimental Results Previous: Experimental Results

Selecting the Client Workloads

**Figure 5:** **Web server load profile:** Average response time in ms, per KB of the request, as a function of the average client arrival rate at the server in our topology (Figure 4(b)).
$\begin{figure}\begin{center} \epsfig{file=final_version_figs/load-rate.eps,width=3in}\end{center}\vspace{-0.2in} \vspace{-0.1in} \end{figure}$

In Figure 5 we show the average response time per KB of client requests (i.e., the completion time for a request divided by the size of the request in KB), as a function of the average arrival rate of clients at the server (i.e., $\frac{100}{\lambda}$ requests/s). The response time quickly degrades beyond an arrival rate of about 15 requests/s beyond which it increases only marginally with the request rate. We select five different points on this load curve (highlighted), corresponding to arrival rates of 1.7, 3.3, 10, 13.3 and 20 requests/s , and evaluate the proposed schemes under these workloads. These workloads represent various stress levels on the server , while also ensuring that it is not overloaded. The high variability in response times in overload regimes might impact the confidence or accuracy of our comparison of the proposed schemes.

In the remainder of the evaluation we focus on addressing the following questions:

To what extent do the route control schemes improve the performance of the multihomed site, relative to using the single best provider alone?
Does employing historical samples help in better estimating future provider performance?
How do active and passive measurement schemes compare in terms of the performance improvement they offer? Which of the two active measurement schemes - SlidingWindow or FrequencyCounts - works better?
At what time intervals should samples for provider performance be collected?
What overheads do the proposed mechanisms incur?

**Figure:** **Performance improvement:** The performance metric $\mathcal{R}$ for the passive measurement scheme with EWMA parameter $\alpha = 0$ (no history employed) and sampling interval of 30s. The graph also shows the performance from the three individual providers.
$\begin{figure}\begin{center} \epsfig{file=final_version_figs/itworks.eps,width=3in}\end{center}\vspace{-0.2in} \vspace{-0.1in} \end{figure}$

**Figure 7:** **Unrolling the averages:** Ratio and the difference in the response times from using just ISP 3 for all transfers relative to using the passive measurement scheme. The average client arrival rate in either case is 13.3 requests/s.
$\begin{figure}\begin{center} \begin{tabular}{cc} \epsfig{file=final_version_fig... ...nce in response times \\ \end{tabular}\end{center}\vspace{-0.2in} \end{figure}$

Next: Improvements from Route Control Up: Experimental Results Previous: Experimental Results

Anees Shaikh 2004-05-05