|Red Hat Enterprise Linux 4: Introduction to System Administration
|Chapter 5. Managing Storage
Hard drive performance characteristics have already been introduced in Section 4.2.4 Hard Drives; this section discusses the matter in more depth. This is important for system administrators to understand, because without at least basic knowledge of how hard drives operate, it is possible to unwittingly making changes to your system configuration that could negatively impact its performance.
The time it takes for a hard drive to respond to and complete an I/O request is dependent on two things:
The hard drive's mechanical and electrical limitations
The I/O load imposed by the system
The following sections explore these aspects of hard drive performance in more depth.
Because hard drives are electro-mechanical devices, they are subject to various limitations on their speed and performance. Every I/O request requires the various components of the drive to work together to satisfy the request. Because each of these components have different performance characteristics, the overall performance of the hard drive is determined by the sum of the performance of the individual components.
However, the electronic components are at least an order of magnitude faster than the mechanical components. Therefore, it is the mechanical components that have the greatest impact on overall hard drive performance.
The most effective way to improve hard drive performance is to reduce the drive's mechanical activity as much as possible.
The average access time of a typical hard drive is roughly 8.5 milliseconds. The following sections break this figure down in more detail, showing how each component impacts the hard drive's overall performance.
All hard drives produced today have sophisticated embedded computer systems controlling their operation. These computer systems perform the following tasks:
Interacting with the outside world via hard drive's interface
Controlling the operation of the rest of the hard drive's components, recovering from any error conditions that might arise
Processing the raw data read from and written to the actual storage media
Even though the microprocessors used in hard drives are relatively powerful, the tasks assigned to them take time to perform. On average, this time is in the range of .003 milliseconds.
The hard drive's read/write heads only work when the disk platters over which they "fly" are spinning. Because it is the movement of the media under the heads that allows the data to be read or written, the time that it takes for media containing the desired sector to pass completely underneath the head is the sole determinant of the head's contribution to total access time. This averages .0086 milliseconds for a 10,000 RPM drive with 700 sectors per track.
Because a hard drive's disk platters are continuously spinning, when the I/O request arrives it is highly unlikely that the platter will be at exactly the right point in its rotation necessary to access the desired sector. Therefore, even if the rest of the drive is ready to access that sector, it is necessary for everything to wait while the platter rotates, bringing the desired sector into position under the read/write head.
This is the reason why higher-performance hard drives typically rotate their disk platters at higher speeds. Today, speeds of 15,000 RPM are reserved for the highest-performing drives, while 5,400 RPM is considered adequate only for entry-level drives. This averages approximately 3 milliseconds for a 10,000 RPM drive.
If there is one component in hard drives that can be considered its Achilles' Heel, it is the access arm. The reason for this is that the access arm must move very quickly and accurately over relatively long distances. In addition, the access arm movement is not continuous — it must rapidly accelerate as it approaches the desired cylinder and then just as rapidly decelerate to avoid overshooting. Therefore, the access arm must be strong (to survive the violent forces caused by the need for quick movement) but also light (so that there is less mass to accelerate/decelerate).
Achieving these conflicting goals is difficult, a fact that is shown by how relatively much time the access arm movement takes when compared to the time taken by the other components. Therefore, the movement of the access arm is the primary determinant of a hard drive's overall performance, averaging 5.5 milliseconds.
The other thing that controls hard drive performance is the I/O load to which a hard drive is subjected. Some of the specific aspects of the I/O load are:
The amount of reads versus writes
The number of current readers/writers
The locality of reads/writes
These are discussed in more detail in the following sections.
For the average hard drive using magnetic media for data storage, the number of read I/O operations versus the number of write I/O operations is not of much concern, as reading and writing data take the same amount of time. However, other mass storage technologies take different amounts of time to process reads and writes.
The impact of this is that devices that take longer to process write I/O operations (for example) are able to handle fewer write I/Os than read I/Os. Looked at another way, a write I/O consumes more of the device's ability to process I/O requests than does a read I/O.
A hard drive that processes I/O requests from multiple sources experiences a different load than a hard drive that services I/O requests from only one source. The main reason for this is due to the fact that multiple I/O requesters have the potential to bring higher I/O loads to bear on a hard drive than a single I/O requester.
This is because the I/O requester must perform some amount of processing before an I/O can take place. After all, the requester must determine the nature of the I/O request before it can be performed. Because the processing necessary to make this determination takes time, there is an upper limit on the I/O load that any one requester can generate — only a faster CPU can raise it. This limitation becomes more pronounced if the requester requires human input before performing an I/O.
However, with multiple requesters, higher I/O loads may be sustained. As long as sufficient CPU power is available to support the processing necessary to generate the I/O requests, adding more I/O requesters increases the resulting I/O load.
However, there is another aspect to this that also has a bearing on the resulting I/O load. This is discussed in the following section.
Although not strictly constrained to a multi-requester environment, this aspect of hard drive performance does tend to show itself more in such an environment. The issue is whether the I/O requests being made of a hard drive are for data that is physically close to other data that is also being requested.
The reason why this is important becomes apparent if the electromechanical nature of the hard drive is kept in mind. The slowest component of any hard drive is the access arm. Therefore, if the data being accessed by the incoming I/O requests requires no movement of the access arm, the hard drive is able to service many more I/O requests than if the data being accessed was spread over the entire drive, requiring extensive access arm movement.
This can be illustrated by looking at hard drive performance specifications. These specifications often include adjacent cylinder seek times (where the access arm is moved a small amount — only to the next cylinder), and full-stroke seek times (where the access arm moves from the very first cylinder to the very last one). For example, here are the seek times for a high-performance hard drive:
Actually, this is not entirely true. All hard drives include some amount of on-board cache memory that is used to improve read performance. However, any I/O request to read data must eventually be satisfied by physically reading the data from the storage medium. This means that, while cache may alleviate read I/O performance problems, it can never totally eliminate the time required to physically read the data from the storage medium.
Some optical disk drives exhibit this behavior, due to the physical constraints of the technologies used to implement optical data storage.