Disk Scheduling in Operating System

Operating System / Tuesday, July 2nd, 2019

The disk is a resource which has to be shared. It therefore has to be scheduled for use, according to some kind of scheduling system. The secondary storage media structure is one of the vital parts of the file system. Disks are the one, providing a lot of the secondary storage. As compared to magnetic tapes, disks have very fast access time and disk bandwidth. The access time has two major constituents: seek time and rotational latency.

The seek time is the time required for the disk arm to move the head to the cylinder with the desired sector. The rotational latency is the time required for the disk to rotate the desired sector until it is under the read-write head. The disk bandwidth is the total number of bytes transferred per unit time.

Both the access time and the bandwidth can be improved by efficient disk I/O requests scheduling. Disk drivers are large single dimensional arrays of logical blocks to be transferred. Because of the large usage of disks, proper scheduling algorithms are required.

A scheduling policy should attempt to maximize throughput (defined as the number of requests serviced per unit time) and also to minimize mean response time (i.e., average waiting time plus service time). These scheduling algorithms are discussed below:


FCFS Scheduling

First-Come, First-Served (FCFS) is the basis of this simplest disk scheduling technique. There is no reordering of the queue. Suppose the requests for inputting/ outputting to blocks on the cylinders have arrived, forming the following disk queue:

50, 91, 150, 42, 130, 18, 140, 70, 60

Also, assume that the disk head is initially at cylinder 50 then it moves to 91, then to 150 and so on. The total head movement in this scheme is 610 cylinders, which makes the system slow because of wild swings. Proper scheduling while moving towards a particular direction could decrease this. This will further improve performance. FCFS scheme is clearly depicted in Figure:

FCFS Scheduling

SSTF Scheduling

The basis for this algorithm is Shortest-Seek-Time-First (SSTF) i.e., service all the requests close to the current head position and with minimum seek time from the current head position.

In the previous disk queue sample the cylinder close to critical head position i.e., 50, is 42 cylinder, next closest request is at 60. From there, the closest one is 70, then 91,130,140,150 and then finally 18-cylinder. This scheme has reduced the total head movements to 248 cylinders and hence improved the performance.

Like SJF (Shortest Job First) for CPU scheduling SSTF also suffers from the starvation problem. This is because requests may arrive at any time. Suppose we have the requests in disk queue for cylinders 18 and 150, and while servicing the 18-cylinder request, some other request closest to it arrives and it will be serviced next. This can continue further also making the request at 150-cylinder wait for long. Thus a continual stream of requests near one another could arrive and keep the far away request waiting indefinitely. The SSTF is not the optimal scheduling due to the starvation problem. This whole scheduling is shown in Figure:

SSTF Scheduling

SCAN Scheduling

The disk arm starts at one end of the disk and service all the requests in its way towards the other end, i.e., until it reaches the other end of the disk where the head movement is reversed and continue servicing in this reverse direction. This scanning is done back and forth by the head continuously.

In the example problem, two things must be known before starting the scanning process. Firstly, the initial head position i.e., 50 and then the head movement direction (let it towards 0, starting cylinder). Consider the disk queue again:

91, 150, 42, 130, 18, 140, 70, 60

Starting from 50 it will move towards 0, servicing requests 42 and 18 in between. At cylinder 0 the direction is reversed and the arm will move towards the other end of the disk servicing the requests at 60, 70, 91,130,140 and then finally 150.

As the arm acts as an elevator in a building, the SCAN algorithm is also known as an elevator algorithm sometimes. The limitation of this scheme is that few requests need to wait for a long time because of the reversal of head direction. This scheduling algorithm results in a total head movement of only 200 cylinders. Figure 7 shows this scheme:

SCAN Scheduling

C-SCAN Scheduling

Similar to the SCAN algorithm, C-SCAN also moves the head from one end to the other servicing all the requests in its way. The difference here is that after the head reaches the end it immediately returns to the beginning, skipping all the requests on the return trip. The servicing of the requests is done only along one path. Thus comparatively this scheme gives uniform wait time because cylinders are like circular lists that wrap around from the last cylinder to the first one.


LOOK and C-LOOK Scheduling

These are just improvements of SCAN and C-SCAN but difficult to implement. Here the head moves only till final request in each direction (first and last ones), and immediately reverses direction without going to end of the disk. Before moving towards any direction the requests are looked, avoiding the full-width disk movement by the arm.

The performance and choice of all these scheduling algorithms depend heavily on the number and type of requests and on the nature of disk usage. The file allocation methods like contiguous, linked or indexed, also affect the requests. For example, a contiguously allocated file will generate nearby requests and hence reduce head movements whereas linked or indexed files may generate requests from blocks that are scattered throughout the disk and hence increase the head movements. While searching for files the directories will be frequently accessed, hence the location of directories and also blocks of data in them are also important criteria. All these peculiarities force the disk scheduling algorithms to be written as a separate module of the operating system so that these can easily be replaced. For heavily used disks the SCAN / LOOK algorithms are well suited because they take care of the hardware and access requests in a reasonable order. There is no real danger of starvation, especially in the C-SCAN case. The arrangement of data on a disk plays an important role in deciding the efficiency of data-retrieval.


Leave a Reply

Your email address will not be published. Required fields are marked *