The Method

Figure 4: JIT-seek.

The JIT-seek method is illustrated in Figure 4. The x-axis depicts time, and the two horizontal lines depict a regular IO and an IO with JIT-seek, respectively. With JIT-seek, the read command for an IO operation is delayed and issued just-in-time so that the seek operation takes the disk head directly to the destination block, without incurring any rotational delay at the destination track. Hence, data transfer immediately follows the seek operation. The available rotational slack, before issuing the JIT-seek command, is now preemptible. We can make two key observations about the JIT-seek method. First, an accurate JIT-seek operation reduces the $T_{rot}$ component of the waiting time without any loss in performance. Second, and perhaps more significantly, the ongoing IO request can be serviced as much as possible, or even completely, if sufficient slack is available before the JIT-seek operation for a higher-priority request.

The pre-seek slack made available due to the JIT-seek operation can be used in three possible ways:

• The pree-seek slack can be simply left unused. In this case, a higher-priority request arriving during the slack time can be serviced immediately.

• The slack can be used to perform additional data transfers. Operating systems can perform data prefetching for the current IO beyond the necessary data transfer. We refer to it as free prefetching [13]. Chunking is used for the prefetched data, to reduce the waiting time of a higher-priority request. Free prefetching can increase the disk throughput. We must point out, however, that free prefetching is useful only for sequential data streams where the prefetched data will be consumed within a short time. Operating systems can also perform another background request as proposed elsewhere [13,16].

• The slack can be used to mask the overhead incurred in performing seek-splitting, which we shall discuss next.