U. Glaeser

of D consecutive blocks, one block from each disk, is read into the memory buffer. The number of I/O
steps is reduced by a factor of D over sequentially accessing the file from a single disk. Despite its simplicity,
disk striping is not the best solution for most other data access problems. For instance, generalizing the
above problem to concurrently read N sequential files, a disk-striping solution would read D blocks of a
single file in each I/O. The total buffer space required in this situation is ND blocks. A more resourceefficient
solution is to perform concurrent, independent read I/Os on the different disks. In one parallel
I/O, blocks from D different files are fetched from the D disks; this requires only 1/ Dth
the buffer of a disk
striping solution if the blocks are consumed at the same rates. In fact, if the blocks are consumed at a rate
comparable to the I/O time for a block, then by using independent I/Os only Θ(
D)
blocks of buffer suffice.
In contrast to the uniform access patterns implied by the previous examples, a skewed data access pattern
results in hot spots, in which a single disk is repeatedly accessed in a short time period. The bandwidth
of the multiple-disk system is severely underutilized in this situation, and the performance degrades to
that of a single disk. Consider, for instance, the retrieval of constant data length (CDL) video data, in
which the frames are packed into fixed-size data blocks; the blocks are then placed on the disks using
either striped, random, or other disk allocation policy. If a number of such streams are read concurrently,
the access pattern consists of an interleaving of the blocks that depends on the playback times of the
blocks. For constant-bit rate (CBR) video streams the playback time of a block is fixed, and (assuming
striped allocation) the accesses are spread uniformly across the disks as in the example on multiple file
access. In the case of variable bit rate (VBR) video data streams, the accesses are no longer uniformly
distributed across the disks, but depend on the relative playback times of each of the blocks. Consequently,
both the load on a disk and the load across the disks varies as a function of time. In this case, simply
reading the blocks in the time-ordered interleaving of blocks, may no longer maximize the disk parallelism,
and more sophisticated scheduling strategies are necessary to maximize the number of streams
that can be handled by the I/O system [22].
The abstract model of the I/O system that will be used to analyze the quality of different schedules
is based on the PDM [58]: the I/O system consists of D independent disks, which can be accessed in
parallel, and has a buffer of capacity M,
through which all disk accesses occur. The computation requests
data in blocks—a block is the unit of disk access. The I/O trace of a computation is characterized by a
reference string, which is an ordered sequence of I/O requests made by the computation. In serving a
reference string the buffer manager determines which blocks to fetch and when to fetch them so that the
computation can access the blocks in the order specified by the reference string. The computation waits
for data from the I/O system only when the data are not available in the buffer. Additionally, when an
I/O is initiated on one disk, blocks can be concurrently fetched from other disks. The number of parallel
I/Os that are issued is a measure of performance in this model. Because the buffer is shared by all disks
it is possible to allocate buffer space unevenly to different disks to meet the changing load on different
disks. The PDM assumes unit time I/Os. In many applications like those dealing with streaming data,
data logging or in several scientific computations, where large block sizes are natural, this is a viable and
useful idealization. In these situations, the number of I/Os has a direct relationship to the I/O time. In
other cases where requests are for small amounts of data and access times are dominated by the seek and
rotational latency components, no analytical models are widely applicable. In these cases, empirical
evaluations need to be employed in estimating performance [19,23].
33.4 Mechanisms for Improving I/O Performance
Prefetching and caching are two fundamental techniques that are employed for increasing I/O performance.
Prefetching refers to the process of initiating a read from a disk before the computation demands
the data. In a parallel I/O system, while a read progresses on one disk, reads can be started concurrently
on other disks to prefetch data that are required later. These prefetched blocks are held in the I/O buffer
till needed. In this way a temporary concentration of accesses to a small subset of the disks is tolerated
by using the time to prefetch from the disks that are currently idle; when the locality shifts to the latter
set of disks, the required data are already present in the buffer.
© 2002 by CRC Press LLC