Hardware Reference
In-Depth Information
The aperiodic load was varied across the range of processor utilization unused by
the periodic tasks, and in particular from 3% to 33%. The interarrival times for the
aperiodic tasks were modeled using a Poisson arrival pattern, with average
T
a
, whereas
the aperiodic computation times were modeled using an exponential distribution.
The processor utilization of the servers was set to all the utilization left by the periodic
tasks; that is,
U
s
=1
U
p
. The period of the periodic servers - namely Polling,
DPE, and DSS - was set equal to the average aperiodic interarrival time (
T
a
) and,
consequently, the capacity was set to
C
s
=
T
a
U
s
.
−
In Figure 6.12, the performance of the algorithms is shown as a function of the aperi-
odic load. The load was varied by changing the average aperiodic service time, while
the average interarrival time was set at the value of
T
a
= 100. Note that the data
plotted for each algorithm represent the ratio of the average aperiodic response time
relative to the response time of background service. In this way, an average response
time equivalent to background service has a value of 1
.
0 on the graph. A value less
than 1
.
0 corresponds to an improvement in the average aperiodic response time over
background service. The lower the response time curve lies on these graphs, the better
the algorithm is for improving aperiodic responsiveness.
The EDL server is not reported in the graph since it has basically the same behavior
as IPE for almost any load conditions. In particular, simulations showed that for small
and medium periodic loads the two algorithms do not have significant differences in
their performance. However, even for a high periodic load, the difference is so small
that it can be reasonably considered negligible for any practical application.
Although IPE and EDL have very similar performance, they differ significantly in their
implementation complexity. As mentioned in previous sections, the EDL algorithm
needs to recompute the server parameters quite frequently (namely, when an aperiodic
request enters the system and all previous aperiodics have been completely serviced).
This overhead can be too expensive in terms of CPU time to use the algorithm in
practical applications. On the other hand, in the IPE algorithm the parameters of the
server can be computed off-line, and used at run-time to replenish the server capacity.
As shown in the graph, the TBS and IPE algorithms can provide a significant reduc-
tion in average aperiodic response time compared to background or polling aperiodic
service, whereas the performance of the DPE and DSS algorithms depends on the ape-
riodic load. For low aperiodic load, DPE and DSS perform as well as TBS and IPE,
but as the aperiodic load increases, their performance tends to be similar to that one
shown by the Polling Server.
Search WWH ::
Custom Search