4.2 Micro-benchmarks

To further explore the behavior of the prototype, we use the Intel Iometer benchmark to stress some array configurations in a controlled manner. In all the following micro-benchmarks, we use a seek locality index of 3, as defined in Section 2.3. We measure the throughput in these experiments.

In this experiment, we perform only random read operations on the disk array while maintaining a constant number of outstanding requests. (We examine writes more fully in the next subsection.) The goals are 1) to understand the scalability of the disk array, 2) to understand the behavior of the system under different load conditions, and 3) to validate (part of) the throughput model of Section 2.4.

**Figure 12:** Throughput as a function of array configuration, number of disks, and queue length. The queue lengths are (a) 8, and (b) 32.
$\includegraphics*[width=3in]{eps/thru_new.eps}$

Figure 12 shows that the SR-Array using the RSATF scheduler scales well as we increase the number of disks under this Iometer workload. The RLOOK scheduler is a close approximation of the RSATF scheduler; and the RLOOK-based throughput model closely captures the behavior of the SR-Array, including the throughput degradation experienced when the queue length is short.

The SATF-based striped and RAID-10 systems do not scale as well as the SR-Array. The throughput gap between all these systems, however, narrows as the queue length increases since the SATF scheduler can overcome the lack of rotational replicas when it has a large number of requests to choose from.

We now analyze configurations under foreground write propagation and validate the model in Section 2.4.

Figure 13 shows the throughput results as Iometer maintains a constant queue length of mixed reads and writes. Each write leads to immediate replica propagations; so write ratio and foreground write ratio are the same, namely, 1-p, where p is defined by Equation (8) of Section 2.3.

Among the configurations shown in the figure, RAID-10 has the worst performance under high write ratios. To understand why, consider the propagation of a single write: the $3 \times 2 \times 1$ SR-Array requires a single seek followed by writing 2 rotational replicas in a single cylinder; but a corresponding $3 \times 1 \times 2$ RAID-10 requires 2 seeks so the amount of arm movement tends to be greater.

**Figure 13:** Throughput as a function of foreground write propagation rate and queue length. The total number of disks is six. The queue lengths are (a) 8, and (b) 32.
$\includegraphics*[width=3in]{eps/rw_new.eps}$

The performance of the striped $6 \times 1 \times 1$ configurations degrade slightly for high write ratios as writes are slightly more expensive than reads.

The performance difference between a $3 \times 2 \times 1$ SR-Array and a $6 \times 1 \times 1$ striped system depends on the write ratio with the former better for low write ratios. If we only consider rotational delay, the rotational replication model of Section 2.2 would imply that the cross-over point between them under LOOK/RLOOK scheduling should be close to the 50% write ratio. If we also consider seek distance, the $3 \times 2 \times 1$ SR-Array has worse seek performance so the actual cross-over point is less than 50%.

Because SATF benefits a $6 \times 1 \times 1$ configuration more than RSATF does to a $3 \times 2 \times 1$ configuration, the cross-over point between these systems under SATF/RSATF scheduling is to the left of that under LOOK/RLOOK scheduling. This distance is even greater when the queue is longer (in Figure 13(b)).

The figure also shows that the RLOOK throughput model (Equation (16)) closely tracks the experimental result under varying write ratios.