Lecture 23 of CS537

Book chapter: Section 13.1, 13.2.

Characteristics of Disk Accesses:

The following paper gives a very good introduction to disk access characteristics.
A disk consists of controllers, arms with attached read/write heads, and multiple rotating platters. Each platter has many tracks; a track is a circle on the platter for recording data. A track further consists of sectors, which is typically 512 bytes and is the basic unit for disk read and write. The collection of tracks at the same position on both sides of the platters is called a cylinder group.
Main components of disk read/write latency: seek latency (moving the arms so that the heads are on the right track) , rotational latency (waiting until the right sector is under the read/write head), and transfer latency (reading or writing the data to and from the disk).
In the current technology trend, seek latency and rotational latency are reducing very slowly. What is improving rapidily is the density of bits on a track; that is, more and more bits can be packed into one track. Thus, once the head is on the right sector, reading the data off the disk medium is very fast and the speed keeps improving. In other words, seek and rotational latencies will continue to be the main portion of the disk access latency.
Track buffers: a technique to reduce disk access latency. For details see paper.

Disk scheduling algorithms

(The discussion on disk scheduling was actually at the end of Lecture 25. But logically it belongs here).

First-Come First-Served Scheduling (FCFS): fair, but poor performance (each disk I/O may take a long time);
SCAN scheduling (Elevator algorithm): the disk arms start at the inner most (or the outer most) track of the disk, and move toward the outer most (or the inner most) track of the disk. When the disk arms move to a cylinder the disk services all current requests on that cylinder, after which the arms move to the next cylinder.
Variations on SCAN: C-SCAN, LOOK. (see section 13.2).
Shortest Distance First (Greedy algorithm): the disk always service the next request that will take the disk heads the least amount of time (seek and rotational time together) to reach.
The book talks about Shortest Seek-Time First, which is an early version of the Greedy algorithm that did not consider rotational latency. Rotational latency is typically hard to calculate because finding out exactly on which sector the head is currently at is difficult.

Impact of disk access characteristics on file system design:

Knowing how disk works, and knowing the typical patterns of read and write on file data blocks and file metadata blocks, we can then know how we should allocate file blocks on disks.
For example, if we know that files tend to be accessed sequentially from the first block to the last block, then we should try to allocate all blocks belonging to the same file to be adjacent to each other and such that blocks 0 through N are laid out sequentially in the reverse direction of the disk spinning. So that after the disk head read block 0, it is automatically at block 1.
Example 2: in UNIX, in order to read a file, first the file inode is read, and then the file blocks (whose positions are found out in the inode) are read. Thus, the UNIX file allocation routine typically allocate the inode and the data blocks it points to close to each other (for example, on neighboring tracks).
Example 3: people have found that files under the same directory tend to be accessed close together. Thus, inode and data blocks for files under the same directory should be allocated close to each other on disk, so that the disk head doesn't have to move far when reading those files.
UNIX Fast File System (FFS) uses the concept of cylinder groups to facilitate file allocation. A cylinder group is a collection of neighboring cylinders (a cylinder is the collection of tracks at the same position on the platters). UNIX FFS always try to allocate all inodes of files under the same directory in the same cylinder group, and try to allocate all data blocks of a file to be in the same cylinder group of its inode, and try to allocate all data blocks of a file sequentially on disk in the reverse direction of the disk spinning.
Of course, UNIX FFS will attempt to allocate file blocks according to the above principle. However, if the blocks in the desire positions are already used, UNIX FFS will find other locations on disk.