Information Technology Reference
In-Depth Information
The CEA Curie and RICC task workloads are available at the Parallel Work-
When analyzing the EET of the tasks in the studied real-world workloads, we
found that most EETs are within either less than 20% or more than 80% of the
maximum execution time allowed in the system. Hence, we propose grouping
tasks in the workloads in 5 different groups: in the first group tasks which have
EET between 0% and 20% of the maximum execution time, in the second groups
tasks with EET between 20% and 40%, then between 40% and 60%, then between
60% and 80%, and finally between 80% and 100% (see Fig. 1).
When averaging the results for the three real-world workloads, we see that in
average 50% of the tasks request less than 20% of the maximum allowed execu-
tion time, 45% of the tasks request more than 80% of the maximum execution
time, and the remaining 5% is somewhat uniformly distributed.
Regarding
ʔ
ET
, the workload analysis showed that the estimation errors are
rather large and, again, not uniformly distributed. Further analysis showed that
a significantly large number of tasks present either a quite accurate estimation
or a very inaccurate estimation. This is shown in Fig. 1. This empirical findings
are similar to the ones presented by Tsafrir [19]. Based on this data we propose
three different error scenarios for our model:
ʔ
low
ET
with an average error of 48%,
with an average error of 56%, and
ʔ
high
ET
ʔ
med
ET
with an average error of 67%.
relative EET percentage of tasks
[0
,
20%)
percentage of tasks
ʔ
low
relative EET error
ET
ʔ
med
ʔ
high
ET
50%
ET
[20%
,
40%)
2%
[0
,
40%)
45% 35%
25%
[40%
,
60%)
2%
[40%
,
75%)
30% 25%
15%
[60%
,
80%)
1%
[75%
,
95%)
5% 10%
20%
[80%
,
100%]
45%
[95%
,
100%]
20% 30%
40%
60%
100
CEA Curie
RICC
Cluster FING
Curie
RICC
Cluster FING
90
50%
80
70
40%
60
30%
50
40
20%
30
20
10%
10
0%
0
[0%,40%)
[40%,75%)
[75%,95%)
[95%,100%]
[0%,20%)
[20%,40%)
[40%,60%)
[60%,80%Ϳ [80%,100%]
Relative execution time error
Tasks estimated execution time
Fig. 1.
Analysis of the proposed workloads
5.2 The Energy Consumption Uncertainty Model
We conducted a set of empirical evaluations in order to determine the uncertainty
model for the energy consumption.
Our starting point was the high-level theoretical linear increasing model that
we originally introduced in our previous work [16]. This model proposes a linear