3.1 Overview

To show the feasibility of our approach, we have developed an infrastructure for experimenting with the various techniques. Figure 4 illustrates the system components and how they stack against each other. There are a number of ways of configuring the system. The controlling agent (at the top) can be either a user level disk driver or an OS kernel device driver. The devices (at the bottom) can be either a disk simulator or real SCSI disks. The remaining components (in the middle) are shared across all configurations and are built in a layered fashion.

The SCSI Abstraction Layer abstracts device specific operations such as SCSI device detection at boot time, issuing SCSI commands, and retrieval of command status. We currently support two different 10000 RPM SCSI drives: Seagate ST34502LW and ST39133LWV; but all experimental results reported are based on the ST39133LWV.

The Calibration Layer is used for calibrating the disk and extracting information regarding the physical layout of the disk. It keeps track of where the disk head is currently located and calculates how much time is required to move the head from its current position to a target sector. Section 3.2 provides more details about our head-tracking technique.

  

Figure 4: Prototype architecture.

A parallel layer to the SCSI abstraction layer is the Simulator. We decided to integrate the simulator into the architecture to shorten the simulation time for long traces: we not only eliminate idle time, but also replace I/O time (which can be long) with simulated time. The simulator also provides the flexibility of exploring the impact of changing disk characteristics. To faithfully simulate the behavior of the disks that we currently use in the prototype, the simulator receives timing information from the calibration layer to configure itself.

The Scheduling layer implements several disk scheduling policies. This layer maintains a read/write command queue for each physical disk and invokes a user determined policy to pick the next request at each scheduling step. We call these queues the drive queues. Sections 3.3 and 3.4 provide details of the scheduling policies.

The Disk Configuration Layer provides support for configuring a collection of disks using techniques such as D-way mirroring, striping, SR-Array, and SR-Mirror. It translates I/O requests for a logical disk to a set of I/O commands on the physical disks and inserts them into the appropriate drive queues. The striping unit is 64K bytes in our experiments.

The topmost Logical Disk Layer is in charge of exposing the logical disk to the application. The kernel device driver exposes a mount point (e.g. drive letter Z: in Windows 2000) to user space. The user level driver exposes the disk in the form of an API to the application.