Information Technology Reference
In-Depth Information
to fully understand all the required steps to deploy a script using different under-
lying applications. More complex scripts can be extended using standard Python
programming techniques.
4.2 Standard SISIM Scalability Test
In the first test the simulation workload was distributed using our strategy,
without any code modifications to the existing
sisim
program. We implemented
a Python script that set the parameters and data files pointing to
sisim
exe-
cutable. Also a simple
main
application was developed. This
main
application
creates a Task that will configure the target code input parameters, uploads the
target code and corresponding input data to the shared storage system and fi-
nally send the Task for execution. The code was instrumented to report partial
and overall execution times.
This case study includes a comparison between the base execution of standard
sisim
routine and the distributed version up to 32 cores using the homogeneous
cluster. Timing and speedup results can be viewed in Table 1. Base case denotes
an execution of
sisim
with
N
= 96 (number of simulations) and a domain
ʩ
of
2880000 points. Single node tests use the strategy to distribute simulations in
one machine as independent native system processes, namely, up to four parallel
workers running 24 simulations each one. Distributed tests use four workers
in each node, using the implemented round-robin scheduler to assign workload.
Under this configuration,
sisim
was parallelized up to 32 instances, each Worker
running 3 simulations.
Table 1.
Standard algorithm parallelized and distributed
case
processes time[s] speedup eciency
base
1
9124.64
1.00
1
9240.11
0.99
99%
2
4748.68
1.91
95%
sisim
single node
3
3240.30
2.82
94%
4
2518.77
3.62
91%
4
2890.09
3.16
79%
8
1515.58
6.02
75%
sisim
distributed
12
1101.01
8.29
69%
16
909.53
10.03
63%
24
712.66
12.80
53%
32
627.36
14.54
45%
Table 1 shows that using the proposed strategy it is possible to increase the
speedup in single node, having an eciency of 91% when using all available re-
sources on a local workstation. Distributed tests results also demonstrate scala-
bility reducing overall computation time progressively. However, it shows that we
lose eciency when more nodes were added. This phenomenon is explained by
