Green Web Services:

joeneetscompetitiveSecurity

Nov 3, 2013 (3 years and 1 month ago)

64 views

Green
Web Services:

Improving
Energy Efficiency in
Data Centers via Workload
Predictions


Massimiliano
Menarini,
Filippo Seracini
, Xiang Zhang,
Tajana

Rosing
, Ingolf
Krüger


GREENS 2013


Our Message:

A Top
-
down Approach


The application layer contains fundamental
information on the execution of a workflow


There are useful correlations between service calls
that we can leverage to optimize the overall behavior
of the system


Leverage that information to predict
future levels of workload and
proactively allocate resources

S.O.PR.A Methodology

Statistical
Model
Historical
usage
trends
Dynamic
Proactive

Allocation
·

Resources
allocated only
when needed
·

Save energy
·

More job
throughput
Execution Model
S
2
S
1
Workload
Predictions
Real Time
Workload
Measurements
S
2
S
1
[
40
%]
[
60
%]
{
5
min
+/
-

30
s
}
Performance
Model
LEGEND
Partially utilized
server
Server OFF
Idle server
(
150
-
200
W
)
Fully utilized server
(
300
-
350
W
)

Accuracy of the workload predictions


A small standard deviation of the time dependency is key to
save more energy


Faced Issues

0
1000
2000
3000
4000
5000
6000
30
210
390
570
750
930
1110
1290
1470
1650
1830
2010
2190
2370
2550
2730
2910
3090
3270
3450
3630
3810
Response Time (ms)

Time (s)

Static 70%
SOPRA 70%
SOPRA 80%
Static 80%
Open Questions for Further
D
iscussion


Cross
-
layer monitoring, modeling and
prediction


How the different layers (application, middleware,
OS, VM, PM) affect resource usage?


How can we model those layers and their
interactions so to take into account also resource
contention?


How can we measure, and what to measure, at
each layer without affecting performance?