The Read Case

Disk drives are optimized for sequential access, and they continue prefetching data into the disk cache even after a read operation is completed [17]. Chunking for a read IO requests is illustrated in Figure 2. The x-axis shows time, and the two horizontal time lines depict the activity on the IO bus and the disk head, respectively. Employing chunking, a large $T_{transfer}$ is divided into smaller chunk transfers issued in succession. The first read command issued on the IO bus is for the first chunk. Due to the prefetching mechanism, all chunk transfers following the first one are serviced from the disk cache rather than the disk media. Thus, the data transfers on the IO bus (the small dark bars shown on the IO bus line in the figure) and the data transfer into the disk cache (the dark shaded bar on the disk-head line in the figure) occur concurrently. The disk head continuously transfers data after the first read command, thereby fully utilizing the disk throughput.

Figure 2: Virtual preemption of the data transfer.

Figure 3 illustrates the effect of the chunk size on the disk throughput using a mock disk. The optimal chunk size lies between $a$ and $b$. A smaller chunk size reduces the waiting time for a higher-priority request. Hence, Semi-preemptible IO uses a chunk size close to but larger than $a$. For chunk sizes smaller than $a$, due to the overhead associated with issuing a disk command, the IO bus is a bottleneck. Point $b$ in Figure 3 denotes the point beyond which the performance of the cache may be sub-optimal³.

Figure 3: Effect of chunk size on disk throughput.