Information Technology Reference
In-Depth Information
freed buffers cannot be reused
.
.
.
. . .
Transmission
. . .
. . .
. . .
. . .
. . .
Disk 0
. . .
. . .
. . .
Disk 1
. . .
. . .
. . .
Disk 2
unnecessary buffer holding time
kd 2 buffers
kd 2 buffers
Figure 3.12 Buffer management under concurrent schedule
which are used to store the retrieved data and the other half storing the data for transmission
as shown in Figure 3.12.
Now it becomes evident that this concurrent schedule does not scale up well. For example,
with Q
20, the per disk buffer requirement for 1-disk and 8-disk arrays
are 2.5MB and 20MB respectively. This gets worse as the array grows larger. Again, buffer
requirement may not be the limiting factor as memory cost has decreased dramatically over the
years. The real limiting factor is the start-up delay, which equals 1.5 times the service round
length. Thus for concurrent schedule the start-up delay increases linearly with the number of
disks in the array (i.e., dT avg ).
=
64KB and k
=
3.5.3.2 Offset Schedule
As illustrated in Figure 3.12, part of the reason for the increased buffer requirement is the
unnecessary buffer holding time after a data block is completely transmitted. The empty buffers
cannot be reused until the next disk service round starts. To eliminate this deficiency, we can
offset the schedules of the disks' service rounds by T a v g seconds as depicted in Figure 3.13
so that a new disk service round is started once every T a v g seconds, aligned with the instants
when empty buffers become available.
This offset schedule has the same buffer requirement for disk retrieval ( d 2 kQ bytes) but the
buffer requirement for transmission is reduced from d 2 kQ bytes to dkQ bytes. Therefore, the
total buffer requirement is reduced from 2 d 2 kQ to ( d
+
1) dkQ bytes. Using the same example
in the previous section ( Q
=
64KB, k
=
20) the per-disk buffer requirement is reduced from
20MB to 11.25MB for an 8-disk array.
3.5.3.3 Split Schedule
Despite the reduction, the per-disk buffer requirement under offset schedule still increases with
the number of disks in the array. The ultimate solution is to divide the streams into d groups and
serve them in separate service rounds of shorter duration - split schedule , shown in Figure 3.14.
Search WWH ::




Custom Search