Experimental Setup

Table 1 shows the factors in the $\langle \vec{W}, \vec{R}, \vec{C}\rangle$ vectors for a storage server. We benchmark an NFS server to evaluate our methodology. In our evaluation, the factors in $\vec{W}$ consist of samples that yield four types of workloads: SPECsfs97, Web server, Mail server, and DB_TP (Section 5.2). The controller uses Fstress to generate samples of $\vec{W}$ that correspond to these workloads. We report results for a single factor in $\vec{R}$: the number of disks attached to the NFS server in $\langle 1, 2, 3, 4 \rangle$, and a single factor in $\vec{C}$: the number of nfsd daemons for the NFS server chosen from $\langle 1,2,4,8,16,32,64,100 \rangle$ to give us a total of $32$ samples.

The workbench tools can generate both virtual and physical machine configurations automatically. In our evaluation we use physical machines that have $800$ MB memory, $2.4$ GHz x86 CPU, and run the $2.6.18$ Linux kernel. To conduct an experiment, the workbench controller first prepares an experiment by generating a sample in $\langle \vec{W}, \vec{R}, \vec{C}\rangle$. It then consults the benchmarking policy(ies) in Sections 3.4-4 to plot a response surface and/or search for the peak rate for a given sample with target confidence and accuracy.