Information Technology Reference
In-Depth Information
Table 1.
Speedup and eciency achieved with the different configurations
SRS
PRS
e
ciency
3N16T Taskmasters Balanced
3.88
4.57
0.2425
3N20T Taskmasters Balanced
3.88
4.56
0.1940
3N12T Taskmasters Balanced
3.66
4.31
0.3050
2N20T Taskmasters Balanced
3.65
4.29
0.1825
2N8T Taskmasters Unbalanced
3.60
4.24
0.4500
2N8T Taskmasters Balanced
3.59
4.23
0.4487
3N24T Taskmasters Balanced
3.24
3.82
0.1350
2N8T Balanced
2.68
3.16
0.3350
2N20T Taskmasters Unbalanced
2.68
3.15
0.1340
1N4T Unbalanced
2.59
3.05
0.6475
3N20T Taskmasters Unbalanced
2.41
2.84
0.1205
1N4T Balanced
2.40
2.82
0.6000
3N24T Taskmasters Unbalanced
2.31
2.72
0.0962
3N16T Taskmasters Unbalanced
2.11
2.49
0.1318
2N20T Balanced
2.07
2.43
0.1035
3N24 Balanced
1.91
2.25
0.0796
3N12T Taskmasters Unbalanced
1.90
2.24
0.1583
2N8T Unbalanced
1.61
1.90
0.2000
3N20T Unbalanced
1.20
1.41
0.0600
3N16T Unbalanced
1.19
1.41
0.0743
3N12T Unbalanced
1.14
1.34
0.0949
2N20T Unbalanced
1.12
1.32
0.0560
3N24T Unbalanced
1.07
1.26
0.0445
Table 1 reports the SRS, PRS and eciency metrics of the different configu-
rations based on Fig. 8.
The comparison between the sequential version of the yaw angle estimation al-
gorithm against the best parallel implementation (i.e., 3N16T with Taskmasters
and load balance) reported in Table 1 demonstrates that approximately 4
of
speedup can be achieved. In terms of eciency, it can be seen comparing all the
unbalanced configurations with and without taskmasters, that using the lasts the
performance was increased at least 2
×
. Obviously, the best eciency (0.64) was
achieved avoiding the use of the network (i.e., the configuration 1N4T which has
only one node and no load balance). Nevertheless, this configuration is not able
to ensure enough speed to process the images in less than 0
.
1s to close which is a
constraint to use the yaw angle control loop. This speed could be accomplished
only with the configurations that uses more than one node like 3N16T, 3N20T,
3N12T, 2N20T, 2N8T and 3N24T using Master/Taskmasters/Slaves hierarchy.
Similar results are obtained when processing the QA3 data set. They are
presented in Fig. 9. Given that this data set has smaller images, the processing
time is slightly lower.
×
