Information Technology Reference
In-Depth Information
Figure 3. Simulation results - mean job completion time
not use any file related information (Rnd, QL).
As soon as file sizes are also considered (Static
DF) the mean job time of all jobs on the Grid is
about 40-60% lower than in the previous case.
Due to the more sophisticated file transferring
approach, using the dynamic data feeder sched-
uler leads to even better (about 5-20% lower)
mean job times. Besides, compared to the QL
scheduler, the jobs scheduled by the dynamic
strategy are finished 40-70% sooner. However
the difference can further increase as CE queues
would enlarge.
Another set of simulations was carried out for
the Static Data Feeder strategy mainly for under-
lining the importance of refined job descriptions
(see Figure 4). These measurements had been
configured in such a way that the jobs monitored
were consuming 10 times more CPU for the second
execution than during their first run.
Four cases were compared: in the case of
Static DF 1 the real execution of our jobs took 10
times longer than the values the scheduler was
using during its calculations. There is an up to
4-5% speedup with the Static DF 2 strategy, which
uses the real (multiplied) running times of the
jobs during the scheduling process. Static DF 3
uses also the shorter execution time estimates
during scheduling, while the real running times
of the jobs were normal for about 50% of the jobs,
and 10 times more for the other half of the jobs.
Using a merged description (currently a 1-1
weighted average) from the two executions men-
tioned above (Static DF 4 strategy) will also reduce
the mean job times with about 10% compared to
the previous strategy.
Refining further these job descriptions with
the execution of the monitored jobs can increase
the credibility of the scheduling strategy, resulting
Search WWH ::
Custom Search