Check out the new USENIX Web site. next up previous
Next: Experimental Results Up: Implementation Previous: Updating and Recovering the


Validating the Integrated Simulator


Table 1: Platform characteristics.
Operating system Microsoft Windows 2000
CPU type Intel Pentium III 733 MHz
Memory 128 MB
SCSI Interface Adaptec 39160
SCSI bus speed 160 MB/s
Disk model Seagate ST39133LWV 9.1 GB
RPM 10000
Average seek 5.2 m$s$ read, 6.0 m$s$ write


Figure 8: Comparison of response times on the EW-Array prototype and those predicted by the simulator as we vary the read/write ratio.
\begin{figure}
\centerline {\psfig{figure=figs/real_response.eps,width=3.1in,clip=}}
\end{figure}

Figure 9: Comparison of throughput on the EW-Array prototype and that predicted by the simulator as we vary the per-disk queue length.
\begin{figure}
\centerline {\psfig{figure=figs/real_throughput.eps,width=3.1in,clip=}}
\end{figure}

Due to the large number of configurations and the long traces that we must experiment with, the experimental results reported in Section 6 are based on those obtained on the simulator; therefore it is necessary to validate the EW-Array simulator using our EW-Array prototype. Table 1 lists some of the platform characteristics of the prototype. We run a benchmark called ``Iometer'', a benchmark developed by the Intel Server Architecture Lab [15]. Iometer can generate workloads of various characteristics including read/write ratio, request size, and the maximum number of outstanding requests. Figure 8 compares the response times measured on a number of six-disk EW-Array prototype configurations with those predicted by the simulator as we vary the read/write ratio. In these Iometer experiments, the number of outstanding requests is one, and the dilution factor of the EW-Array is two. The two EW-Array configurations ( $D_m \times D_s
\times D_d = 1 \times 3 \times 2$, and $D_m \times D_s \times D_d = 2
\times 1.5 \times 2$) have similar response times and they are closely matched by those predicted by the simulations. Since the eager-writes have much lower latency than reads, the response time decreases as the write ratio increases. Figure 9 compares the throughput obtained on the same six-disk EW-Array configurations with that predicted by the simulator as we vary the queue length per disk. In these Iometer experiments, the write ratio is 50%. As the per-disk queue length increases, the $1 \times 4.8 \times 1.25$ EW-Array achieves greater throughput than the $2 \times 1.8 \times 1.6$ configuration because it becomes increasingly difficult for the latter configuration to mask the replica propagation even with a larger dilution factor. The throughput measured on the prototype matches closely the simulated result.
next up previous
Next: Experimental Results Up: Implementation Previous: Updating and Recovering the
Chi Zhang
2001-11-16