Tuning the Storage Subsystem - Expert Oracle RAC Performance Diagnostics and Tuning - page 352

As seen in this formula, the access time is affected by the seek time and by the rotational delay (measured as
latency) of the disks. This means the access time is the sum of the seek time and rotational latency. So, for a disk with
4 millisecond (ms) seek time and 4ms rotational latency, it could take about 8ms from the moment the disk initiates
the read request (seeks to find the data) to the moment when it actually starts reading it. The advent of faster disks,
which essentially means disks spin faster (or have a higher RPM), translates into a reduced rotational delay.
However, by the same measure, the seek time has not reduced very significantly over the years. Let's delve a little
deeper into this seek time.
The seek time of a hard disk measures the amount of time required for the read/write heads to move between
tracks over the surfaces of the platters. Seek time is one of the most important metrics when discussing performance
specifications. Switching between tracks requires the head actuator to move the head arms physically. Head arm
movement between tracks is a mechanical operation and takes a specific amount of time. This time required depends
on the distance between the tracks. As in most mechanical operations, there is certain overhead involved in track
switching.
Seek time is normally expressed in milliseconds, with average seek times for most drives today ranging from 5ms
to 7ms. When computing the overall response time for a user request or disk operation, a millisecond is an enormous
amount of time. Considering that the speed of the system bus is measured in nanoseconds, even small reductions in
seek times can result in improvements in overall system performance.
There are three different seek-time specifications defined by disk drive manufacturers:
• Average seek time: It is the average time to seek the data from one random track (cylinder) to
any other.
• Track-to-track: This is the time required to seek between adjacent tracks, is usually
around .4ms.
• Full stroke: This is the amount of time to seek the entire width of the disk, from the innermost
track to the outermost. The typical range is between 10.2ms and 12ms.
Compare seeking from adjacent tracks (track-to-track) with seeking end to end (full stroke). At first glance it is
obvious that it takes longer for the head to move to the next track as compared to the head moving across several
tracks. However, when examined closely, the difference is only 3%. Where is all the time taken? A significant portion of
the seek time is really settling time, which is a pre-determined delay programmed in the drive electronics.
The database applications have always favored several disks of smaller capacity, because the number of disks,
and not the overall storage capacity of those disks, determines the limit on I/O operations. More disks meant more
read/write heads, which in turn meant higher number of I/O operations, which helped in the overall performance
of the database. Despite the I/O advantages obtained by using many disks of small capacity, unfortunately
manufacturers these days only make large-capacity disks. To complicate matters, these days there are choices within
and across drive families; for example, 73GB 15k RPM, 73GB 10K RPM, 146GB 15K RPM, 300GB 10K RPM, or
750GB 10K RPM.
Based on the above observation, it makes sense to have fewer expensive 10K disks or even 7200RPM disks as
compared to a few expensive 15K RPM disks. One should size for the number of disk read/write heads as opposed to
capacity or even rotational speed.
Moreover, as the drive capacity increases within a particular family of drives, the performance per physical drive
may not increase. For example, when comparing a 146GB 10K RPM drive with a 300GB drive, though the capacity
doubles, the access density , or the number of I/Os per gigabyte that the drive can perform, is cut in half. Thus, it is
important to strike a balance between performance and cost by configuring the system based on the number of drives
required for performance rather than to configure them for the required capacity.