The I/O subsystem is a series of processes responsible for moving blocks of data between disk and memory. In general, each task performed by either kernel or user consists of a utility performing any of the following (or combination thereof):
Reading a block of data from disk, moving it to memory
Writing a new block of data from memory to disk
Read or write requests are transformed into block device requests that go into a queue. The I/O subsystem then batches similar requests that come within a specific time window and processes them all at once. Block device requests are batched together (into an “extended block device request”) when they meet the following criteria:
They are the same type of operation (read or write).
They belong to the same block device (i.e. Read from the same block device, or are written to the same block device.
Each block device has a set maximum number of sectors allowed per request. As such, the extended block device request should not exceed this limit in order for the merge to occur.
The block device requests to be merged immediately follow or precede each other.
Read requests are crucial to system performance because a process cannot commence unless its read request is serviced. This latency directly affects a user's perception of how fast a process takes to finish.
Write requests, on the other hand, are serviced by batch by pdflush kernel threads. Since write requests do not block processes (unlike read requests), they are usually given less priority than read requests.
Read/Write requests can be either sequential or random. The speed of sequential requests is most directly affected by the transfer speed of a disk drive. Random requests, on the other hand, are most directly affected by disk drive seek time.
Sequential read requests can take advantage of read-aheads. Read-ahead assumes that an application reading from disk block X will also next ask to read from disk block X+1, X+2, etc. When the system detects a sequential read, it caches the following disk block ahead in memory, then repeats once the cached disk block is read. This strategy decreases seek time, which ultimately improves application response time. The read-ahead mechanism is turned off once the system detects a non-sequential file access.
