AutoDock Gateway for Molecular Docking Simulations in Cloud Systems - Cloud Computing with e-Science Applications - page 232

1000

Running - DG
Finished - DG
Running - Cloud
Finished - Cloud

800

600

400

200

0

0 20000 40000 60000 80000 100000 120000 140000 160000 180000 200000

Time (seconds)

FIGURE 10.16
Processing Vina inputs in 1,000 jobs.

Figure 10.16 shows the case when the inputs were split into 1,000 jobs.
The x axis of the figure represents the elapsed time, and the y axis repre-
sents the number of (running and finished) jobs. As can be seen, it took
more than twice as long to process the jobs on the DG as on the cloud.
In the case of both the DG and the cloud execution, we can see a short
“running up” period, followed by a steep processing phase; finally, the last
jobs' processing slows.
The job processing figure of the DG case is a bit steeper; thus, processing
jobs on this DCI (with 2,000 to 3,000 active clients) is a bit faster than in
the cloud (with 25 processors). However, in the case of the volunteer DG,
we can clearly observe the tail effect, which means that the last 10%-20%
of the jobs required nearly as long an execution time as the first 80%-90%
of the jobs. The tail effect is missing in the case of the cloud execution;
hence, the overall execution time is much shorter on the cloud. Notice that
the tail effect is a well-known problem of volunteer computing, and there
are several ways of eliminating it [ 12 , 13 ]. One of the possible solutions is
exactly the support of large volunteer DGs with relatively small dedicated
clouds that run the last 10%-20% of the jobs concurrently with the DG.
It has been presented [12] that with such a technique the tail effect can be
reduced significantly.