Energy-Aware Policies in Ubiquitous Computing Facilities - Cloud Computing with e-Science Applications - page 275

VV
nkTq

− V

kT

GS th ds

2

nC W
L

kT
q

/ / q

I = I · e · 1 − e whereI = 2 ··· · ·

leak s s ox

DS > 100 , the contribution of the second exponential is negli-
gible [22], so the previous formula can be rewritten as follows:

When V mV

VV
nkTq

− VV

nkTq

−

GS th GS th

/ 2 /

I = I · e = B Te · ·

leak s

where technology-dependent parameters can be grouped together in a con-
stant B .
Based on the leakage current equation, we describe the leakage power for
a particular machine m as the next equation:

VV
nkTq

−

GS th

2 /

P = I · VB Te = · · · V

leak mleakm , , DD m , DD m ,

As can be seen, leakage has a strong dependence on temperature. Even
though power models have traditionally disregarded leakage, recent studies
are beginning to take it into account. Some cloud computing solutions, such
as those in Reference 23, have considered the dependence of power consump-
tion on temperature due to fan speed as well as the induced leakage current.
Moreover, taking into account the leakage-cooling trade-offs at the server
level by finding an optimum point between the fan power and the leakage
power has proven to yield up to 10% energy savings at the server level [24].
In the case of cloud computing, it is especially interesting to take into
account the temperature of the different computing resources. The pool
of resources that builds the entire cloud infrastructure allows the utiliza-
tion of those resources most appropriate to the operating situation. Thus,
depending on the type of application and the thermal state of the machine,
an efficient allocation can be performed that minimizes the static consump-
tion of the computing infrastructure by keeping the unused resources in a
low-power state.

12.4.2.2 Dynamic Power Modeling

Dynamic power consumption varies depending on the characteristics of
the particular workload to be executed, as well as on the platform where
the workload is executed. The same workload can present different energy
behavior depending on the target platform, as shown in Figure 12.2, obtained
from Reference 25.
To understand and take advantage of these differences, dynamic power has
to be modeled. Dynamic power modeling of enterprise servers has recently
been tackled via the use of performance counters [26, 27]. Performance