Next: Lack of Network Processors
Up: Star, Bus and Trees
Previous: Basic Model
There are at least two ways of returning results in a star network.
Firstly, results can be returned in the inverted order of activating PEs
(Fig.4(a)).
Secondly, they can be returned in the order of activating the PEs
(Fig.4(b)).
Figure 4:
Returning results in the star architecture.
(a) in the inverted order of activating the PEs,
(b) in the order of activating the PEs.
 |
In the first case computing on
lasts as long as
sending data to
, processing on it and returning the
results from
.
Therefore, distribution of the load can be found by
(1) with modified first line:
 |
(6) |
In the second case the time of processing on
and returning results from processor
is equal
to the time of sending to
and processing on
.
Hence, the following modification of the basic equation set
can be used to calculate
's:
Both (6), and (7) can be solved in
time.
However, equations (7)
have a solution even if
is so small
that no schedule of the exact form presented in Fig.4b exists.
To avoid such a case one should verify whether
a non-negative idle time on the network processor of
exists,
i.e. whether the following condition holds:
.
Otherwise, a shorter schedule using fewer processors may be possible.
Next: Lack of Network Processors
Up: Star, Bus and Trees
Previous: Basic Model