Information Technology Reference
In-Depth Information
Fig. 4.
Heat-maps of the
number of task failures
(on top) and the
speedup
(bot-
tom). Results are presented for different combinations of (
i
) the percentage of budget
assigned to spot instances (horizontal axis), and (
ii
) the error on the prediction of the
optimal bid price for the next computation hour (vertical axis).
The largest number of failures (top figure) is circumscribed to scenarios where
the
spotsRatio
parameter (percent of the budget is used for acquiring spot in-
stances) varies from 40% to 100% and the
bidError
parameter (bid price pre-
diction error) ranges from 41.5% to 48.5%. We call this set of scenarios the
max-failures region.
Speedups.
Comparing the top and bottom heat-maps a common pattern is evi-
denced. The speedups achieved in the
max-failures
region are the lowest (bottom
figure). This correlation indicates that the number of failures directly affects the
speedup achieved by SIAA. A larger number of failures not only reduces the
number of instances running but also forces the restart of the tasks affected by
the failure incrementing the workflow makespan.
The best speedups are achieved using a 20%-40% of budget assigned to spot
instances. Within this range, the highest speedup levels correspond to scenarios
where the bid prediction error is below 5%. From the analysis arises that for
achieving the maximum speedup in practice it could be convenient to properly
select the
spotsRatio
parameter considering the quality of the bid price predic-
tion method.
It is also worth to point out that using an infrastructure entirely composed
of on-demand instances (
spotsRatio
= 0) permits the achievement of better
speedups than those corresponding to the max-failures region indicating that
for some
bad quality
bid price predictors it is very important to select the proper
value of
spotsRatio
.
From the results of both experiments, we can conclude that the use of spot
instances plays a fundamental role on the reduction of execution times and costs
of scientific workflows on the cloud. Moreover, SIAA has proved that is a robust
