Check out the new USENIX Web site. next up previous
Next: CPU scheduling Up: Retrofitting Quality of Service Previous: Implementation

Experimental results

  This section demonstrates experimentally that applications can use Eclipse/BSD's /reserv API and CPU, disk, and network schedulers so as to obtain minimum performance guarantees, regardless of other load on the system.

We ran experiments on the configuration shown in Figure 3, where HTTP clients on nodes A to E make requests to the HTTP server on node S. Nodes A to C are Pentium Pro PC's running FreeBSD. Nodes D and E are Sun workstations running Solaris. The operating system varies only in node S, being either FreeBSD or Eclipse/BSD. Node S is a PC with 266 MHz Pentium Pro CPU, 64 MB RAM, and 9 GB Seagate ST39173W fast wide SCSI disk. All nodes are connected by a Lucent P550 Cajun Ethernet switch (unless otherwise noted, at 10 Mbps). Node S runs the Apache 1.3.3 HTTP server and hosts multiple Web sites. Nodes A to E run client applications (some derived from the WebStone benchmark ) that make requests to the server. At most ten clients run at each of the nodes A to E. Unless otherwise noted, all measurements are the averages of three runs.

  
Figure 3: Node S is a Web server that hosts multiple sites on either FreeBSD or Eclipse/BSD.
\begin{figure}
\centerline{\epsfxsize=3in \epsfbox{/home/eclipse/papers-dir/USENIX99-Banu/FIGS/exper.ps}}\end{figure}

Each experiment overloaded one of the server's resources, as described in the following subsections.



 
next up previous
Next: CPU scheduling Up: Retrofitting Quality of Service Previous: Implementation
Jose Brustoloni
4/28/1999