Graphics Programs Reference
In-Depth Information
Mesh size and
processing
idle time
interaction policy
power
2-neigh. pr.
3-neigh. pr.
4-neigh. pr.
2
×
2, preemptive
3.64(90.95%)
9.0%
-
-
2
×
3, preemptive
5.46(90.93%)
9.3%
8.9%
-
3
×
3, preemptive
8.16(90.67%)
9.6%
9.1%
9.2%
2
×
2, not preempt.
1.70(42.62%)
57.4%
-
-
2
×
3, not preempt.
2.44(40.67%)
60.1%
57.8%
-
3
×
3, not preempt.
3.54(39.33%)
58.6%
56.0%
55.6%
Table 11.3: Performance results: idle time and effective processing power
(simplified models)
Mesh size and
processing
idle time
interaction policy
power
2-neigh. pr.
3-neigh. pr.
4-neigh. pr.
2
×
2, preemptive
3.66(91.52%)
8.5%
-
-
2
×
3, preemptive
5.49(91.47%)
8.7%
8.2%
-
3
×
3, preemptive
8.23(91.43%)
9.0%
8.2%
8.4%
2
×
2, not preempt.
2.66(66.39%)
33.6%
-
-
2
×
3, not preempt.
3.96(66.03%)
34.6%
32.7%
-
Table 11.4: Performance results: idle time and effective processing power
•
active executing its local computation
•
active serving a request from one of the neighbours
•
idle, with the local task waiting for service from a neighbour
•
idle, but with the local task being served by a neighbour
Only the active states of the processors are considered in the computation
policies achieve a low throughput if multitasking is not allowed, whereas
preemptive interrupt-like policies are an effective replacement for multitask-
ing. Thus, in this (unfair) comparison the preemptive interaction policy is
dramatically more e
cient than the non-preemptive one.
11.4.2
Mesh models with multitasking
interaction policy when the multitasking behaviour of Fig.
11.5
is adopted
on each processor. The model of the processimg element is sligthly simplified
but the behaviour of the net is the same. Fig.
11.8
shows the model of an
Search WWH ::
Custom Search