JIT-seek: Preempting ${\bf T_{rot}}$

After the reduction of the $T_{transfer}$ component of the waiting time, the rotational delay and seek time components become significant. The rotational period ($T_P$) can be as much as $10$ ms in current-day disk drives. To reduce the rotational delay component ($T_{rot}$) of the waiting time, we propose a just-in-time seek (JIT-seek) technique for IO operations.

Definition 3.2: The JIT-seek technique delays the servicing of the next IO request in such a way that the rotational delay to be incurred is minimized. We refer to the delay between two IO requests, due to JIT-seek, as slack time.

Benefits:

• 1. The slack time between two IO requests is fully preemptible. For example, suppose that an IO request must incur a $T_{rot}$ of $5$ ms, and JIT-seek delays the issuing of the disk command by $4$ ms. The disk is thus idle for $T_{idle}=4$ ms. Then, the expected waiting time is reduced from $2.5$ ms to $\frac{1}{2} \frac{1 \times 1}{1+4} = 0.1$ ms.

• 2. The slack obtained due to JIT-seek can also be used to perform data prefetching for the previous IO or to service a background request, and hence potentially increase the disk throughput.

Overhead: Semi-preemptible IO predicts the rotational delay and seek time between two IO operations in order to perform JIT-seek. If there is an error in prediction, then the penalty for JIT-seek is at most one extra disk rotation and some wasted cache space for unused prefetched data.