Check out the new USENIX Web site. next up previous
Next: Output link scheduling Up: Experimental results Previous: CPU scheduling

Disk scheduling

Again in the second experiment, an increasing number of clients continuously made CGI requests to either of two Web sites hosted at node S. However, these requests are I/O-intensive, consisting of reading a 100 MB file and returning a 10 KB reply. Because requests and replies are small and each request involves considerable disk I/O but little processing, the bottleneck resource is the disk. We reserved 50% of S's disk bandwidth to the Web site of interest and measured the latter's average throughput over three minutes. YFQ's sort queue was configured with a batch size of 4 requests. During the measurements, the site of interest had ten clients and the competing site had a varying number of clients. Figure 6 shows that in the absence of other load, Eclipse/BSD gives to the site of interest essentially all of the bottleneck resource, even though the site has only 50% reserved. When the load on the competing site increases, the throughput of the site of interest decreases. However, on Eclipse/BSD, the throughput bottoms out at roughly the reserved amount, whereas on FreeBSD the throughput decreases without bound. This shows that FreeBSD and Eclipse/BSD are equally good when there is excess disk bandwidth, but when bandwidth is scarce, Eclipse/BSD is also able to guarantee a minimum disk bandwidth allocation.


  
Figure 6: The site of interest gets at least its reserved fraction (50%) of the disk bandwidth.
\begin{figure}
\centerline{\epsfxsize=3.5in \epsfbox{figs/figdiskf.ps}}\end{figure}


next up previous
Next: Output link scheduling Up: Experimental results Previous: CPU scheduling
Jose Brustoloni
4/28/1999