Information Technology Reference
In-Depth Information
Fig. 3.
Performance comparison for each type of workflow. Three autoscaling algo-
rithms are analyzed: (a) Scaling First (ScaF), and (b) Scheduling First (SchF) and (c)
SIAA. Algorithms are compared considering speedup, average execution cost per task
and eciency.
The worst performing method is SchF because the algorithm purchases in-
stances in decreasing cost (and performance) until the hourly budget is con-
sumed. In practice happens that some tasks are scheduled to very slow instances
increasing the overall makespan.
With respect to the
cost of execution
(second row), the graphics show
that SIAA obtains the lowest cost of execution per task (reductions are in the
range 21.5%-43.6%). Again, these results respond to the fact that SIAA is able
to acquire instances with better cost-performance relation. Hence, for the same
levels of performance instances of lower price can be acquired. For SchF and
ScaF the average cost per task is similar. It is noticeable that in general, tasks
for the LIGO and SIPHT workflows are considerably more expensive (about one
order of magnitude). The reason is that such applications comprise tasks of very
long duration (see figure 2).
The third row of graphics present the average
instance percentage of use
for
each algorithm. It can be seen that SIAA presents the lowest usage of instances
(6.6%-16.6% of use below its competitors). The reason for such behavior is that
SIAA acquires a higher number of instances increasing the total
lag time
and
un-
used partial hours
of computation. However, it is worth mentioning that the ma-
jority of this
wasted
instance time corresponds to (cheap) spot instance hours (in
average, SIAA acquired 5.49 spot instances per each on-demand instance).
As a summary, table 2 presents the results of the three strategies considered.
The best results per analyzed aspect (column) are highlighted in bold font.
