Database Reference
In-Depth Information
a much larger logical block size, requiring much larger transaction sizes to
maintain full performance.
Because disks are mechanical devices, there are a number of latencies in-
volved in access. These latencies include: rotational latency, head-switch la-
tency, track-to-track latency, and seek latency. The disk heads are mounted
on mechanical arms that position them over the disk platters. Most disk units
operate the heads on all platters in tandem. After the heads are positioned
over the correct cylinder or track, the disk must wait for the platters to ro-
tate so that the correct sector is located under the read heads. The latency
of the platter rotation is the rotational latency of the disk, and affects the
rate at which the sector can be read or written to the cylinder. Read and
write operations can only engage one head at a time, so the head-switch time
(closely related to the rotational latency of the disk) refers to the amount
of time required to switch between heads on different platters. The latency
of repositioning the head differs greatly depending on whether it is between
neighboring tracks or it involves random accesses. To speed write operations,
many disks employ buffer caches to hide these latencies, but only hide laten-
cies for small transactions. Such buffers cannot mask the reduced throughput
that results from the time spent moving the heads. Consequently, sequential
(append-only writes or streaming reads) offer the best performance for disk
devices. Random access (seeking) presents the worst case read and write per-
formance due to the mechanical latencies involved in repositioning the disk
heads. Solid state nonvolatile storage devices may offer considerable advan-
tages for such access patterns.
A common misconception about disk performance is that the device inter-
face performance defines the expectations for the disk's performance. In fact,
the sustained media transfer performance is often an order of magnitude less
than the performance of the device interface. For example disks that adhere
to the ATA-100 standard boast 100 MB/s transfer rates, but in practice one is
limited by the transfer rate of the raw device, which is typically on the order
of 20 MB/s or less internally. Composite devices such as RAID can share a
common bus (such as SCSI or Fiberchannel) to saturate its available band-
width, and are termed “spindle limited” if the performance of the underlying
disk media is less than the available bandwidth of the device interface.
1.3.2 Performance Characteristics
Although there is significant work at the file system and operating system
levels to hide some of the performance characteristics of block devices, most
of the characteristics of the underlying devices end up percolating up through
the entire software stack and must become first-order design issues for high-
performance scientific I/O applications.
The performance of disk devices is substantially affected by the size of the
I/O transaction. The behavior is partly due to the disk device's organiza-
tion, but also due to the POSIX I/O sequential consistency semantics for
