Gridbus Workflow Enactment Engine - Grid Computing: Infrastructure, Service, and Applications - page 137

100

Waiting
Submitted
Executing
Completed

80

60

40

20

0

0 200 400 600 800 1000 1200

Time (seconds)

FIGURE 5.18 Execution progress for processing 50 subjects.

however, it takes only 45 minutes using i ve grid nodes. The speedup rate
is over 50%. It shows that the performance of conducting fMRI data analy-
sis can be signii cantly improved by using the grid.
Figure 5.18 shows the execution progress for processing 50 subjects. At
the beginning of the workl ow execution, 50 align_warp jobs are generated
for the i rst step, and each job processes one subject image. Once a job in
step 1 is completed, the task manager of step 2 is notii ed by the output
event of this job. It then generates a new reslice job of step 2. Therefore, the
number of waiting jobs does not continuously decrease when align_warp
jobs are completed. In Figure 5.18, we can observe that the number of wait-
ing jobs remains around 50 until 50 jobs of step 1 are completed. All the
results of step 2 are processed once by the softmean task, and the comple-
tion of softmean generates three slicer jobs to produce 2D images along
three dimensions. Therefore, there are only a small number of waiting
jobs after 600 seconds.
Figure 5.19 shows the number of jobs of each task running over the exe-
cution. Available jobs of step 1 and step 2 can be executed in parallel.
However, the jobs of step 1 have a higher priority than those of step 2,
when they compete for resources. As we can observe, most of the execu-
tion jobs during 0-580 seconds are produced by alignWarp and only a few
of the jobs of reslice were executed.
Table 5.4 shows the start and end time for each task in the workl ow for
100 subjects. The start time we measured is the time when the stage-in
of input data to the remote resource is started and the end time is the