Throttling On-Disk Schedulers to Meet Soft-Real-Time Requirements

Abstract

Many contemporary disk drives have built-in queues and schedulers. These features can improve I/O performance, by offloading work from the system's main processor, avoiding disk idle time, and taking advantage of vendor-specific disk characteristics. At the same time, they pose challenges for scheduling requests that have real-time requirements, since the operating system has less visibility and control over service times. While it may be possible for an operating system to obtain more predictable real-time performance by bypassing the on-disk queue and scheduler, the diversity and continuing evolution of disk drives make it difficult to extract the necessary detailed timing characteristics of a specific disk, and to generalize that approach to all hard drives. This paper demonstrates three techniques we developed in the Linux operating system to bound real-time request response times for disks with internal queues and schedulers. The first technique is to use the disk's built-in starvation prevention scheme. The second is to prevent requests from being sent to the disk when real-time requests are waiting to be served. The third is to limit the length of the on-disk queue in addition to the second technique. Our results show the ability to guarantee a wide range of desired response times while still allowing the disk to perform scheduling optimizations. These techniques can be generalized to disks from different vendors.

Coscheduling of CPU and I/O Transactions in COTS-Based Embedded Systems

“…First of all, peripheral activities impose an additional overhead on the CPU: device driver execution. Techniques to account for such overhead have been described in [10,20] for network cards and hard disks based on experimentally-derived bounds. Second, there is another potential source of interference at either the cache or FSB level: other CPUs.…”