"To get data on or off a disk, a drive head has to swing into the correct position and wait for the right part of the disk to come around. This can take hundredths of a second -- many times longer than it takes to access data stored in RAM, for example. As a result, a disk I/O subsystem can be a huge data bottleneck, and significant improvements in the overall performance of the server can be achieved if the effects of this bottleneck can be minimized.
"To understand how this might be done, let's take a look at how data is moved to and from a disk.
"Put simply, the I/O subsystem accepts a stream of requests to read or write data to and from a disk which it holds in a queue. To help speed things up, it usually merges read or write requests together if they are close to each other in the queue, and if they involve the same area of the disk.
"Read requests are generally given higher priority than write requests because a given process will have to wait for the read data it has requested before it can continue, while it won't have to wait for the result of a write.
"The subsystem will also usually detect when data is being read sequentially from the disk and use a technique called "read-ahead" to read and cache the disk block following the one it has been asked to read. This can reduce seek time during sequential reads, although it does nothing to speed up reads to other random parts of the disk, and it is switched off when the subsystem detects random (i.e., non-sequential) reads."
