Information Technology Reference
In-Depth Information
Figure 4. Comparison of smut-fungi dataset on a single CPU and on a Grid using 124 Grid jobs and 150
tasks. Note that there is a rather high redundancy in Grid jobs and not all 124 jobs really participate
in the overall calculation because of start-up latencies. In fact, a few Grid jobs (AxWorker.pl) started,
requested tasks and found out that there were no more tasks available and gracefully finished
Note that the input files are bigger than in the
previous experiment: they cannot be directly
submitted with the Grid job but they are up-
loaded to a Storage Element and then dynami-
cally downloaded by Grid jobs when needed.
A direct performance comparison between the
cluster and a Grid is not feasible since the cluster
we used had several favorable features that a
multi-institutional Grid does not have: a shared
file system between all processing nodes which
minimizes the data transfer time; homogeneous
hardware infrastructure; pre-defined number of
CPUs that are available which does not require
an automatic task assignment, no overhead for
job submission etc. However, the cluster needed
to be reserved in advance (larger slots can only
be obtained overnight) which means that it was
only available at a specific time, whereas Grid
resources are available on demand at any time.
Intuitively, one expects a cluster to provide a bet-
ter response time to a large size application but it
has a considerable “reservation latency”, a fact
that should not be underestimated.
The final performance results of the experi-
ments are depicted in Figure 5. For the Grid ex-
ecution, we used between 90 and 175 parallel jobs
Box 3.
AxParafit.pl -p 100 -1 2048 -2 2048 -3 2025 -4 2031 -A selection_2048.mat \
-B selection_2048_P.pco -C selection_2048_H.tr a -n RUN_2
Search WWH ::




Custom Search