1.7.1 Power consumption models
Modern day’s computer systems are built using the complementary metal-oxide semiconductors (CMOS) technology. There are mainly two components in the power consumption of a CMOS circuit: static anddynamic. The static component is dependent on the system parameters, such as number and type of transistors,
process technology, etc. On the other hand, the dynamic power consumption is basically driven by the circuit activities and traditionally it was believed to be the major contributor [29]. A wide volume of research on scheduling is done which considers only the dynamic power consumption [30, 31, 32, 33, 34].
Mathematically, we write
Ptotal =Pbase+Pdyn (1.1)
where, Pbase is the static power consumption, and Pdyn is the dynamic power consumption.
The dynamic power consumption is expressed as a function of voltage (V) and operating frequency (f). Mathematically, it can be written as follows.
Pdyn = 1
2CV2f (1.2)
Again, there are various power consuming components present in a large system (such as cloud). The total energy (or power) consumption of such a system com- prises of their individual power consumptions. These components include CPU, memory, disk, storage, and network. Various studies [35, 36, 37] have been done to express the total power consumption as a combination of these components.
For instance, Geet al. [35] have used energy consumption of CPU, memory, disk, and NIC (network interface card) to model the energy consumption of the system.
Songet al. [36] expands this model where they expressed the energy consumption as a product of the power consumption of the components with the operation time of the corresponding component.
Though there are many power consuming components in a typical server, the CPU is the major contributor [38, 39] and a wide range of scheduling algorithms are proposed considering the energy consumption of the processor only [30, 40, 41, 42]. As reported in [1], the CPU consumes almost 60% of the total power consumption for a Xeon based server and Figure 1.2 shows the component-wise power consumption for that server. We see that the processor consumes 60% of the total power, and memory is the second highest contributor with 19%. We also present the power breakdown of a server placed at the Google data center using Figure 1.3. Though memory consumes a significant chunk of power, CPU remains to be the major contributor.
Figure 1.2: Component wise power consumption values for a Xeon based server [1]
Figure 1.3: An approximate power breakdown of a server in Google data center [1]
Now the power consumption of a processor is expressed as a function of its uti- lization, processor frequency, etc. Fan et al. [37] made a significant contribution by proposing a power model where the CPU power consumption is expressed as a linear function of its utilization. This can be written using the Equation 1.3.
P(u) =PBase+ (Pbusy−PBase)×u (1.3) where, P(u) is the estimated power consumption at utilization u, PBase is the power consumption when it does not execute any workload (utilization is zero).
This is the static component of the power consumption and for the large systems, it is often considered as 60% to 70% of the total power consumption [43, 44, 45].
Pbusy is the power consumed when the server is fully utilized. A significant amount of research on energy-aware scheduling policies considered this power consumption model in their work [45, 46, 47].
In case of a virtualized environment, the total power consumption can also be expressed as the summation of the individual power consumptions of the VMs and the base power consumption [48], and this can be written as
Ptotal =Pbase+
ν
X
i=1
Pvm(i) (1.4)
where,Pbase indicates the static component, Pvm(i) indicates the power consump- tion of theith VM, andν indicates the total number of VMs placed on the server.
Scheduling policies which target the dynamic power consumption of the hosts (or processors) primarily uses the DVFS (dynamic voltage and frequency scaling) technique by reducing the frequency and voltage to reduce the power consumption.
In DVFS technique, the operating voltage and frequency of the processors are adjusted dynamically to adjust the speed of the processor. This in tern, effects the power consumption of the processor [33, 49, 50, 51, 52]. On the other hand, scheduling policies for the large systems, in general, try to reduce the static power consumption by reducing the number of active components of the system [53, 54, 55, 56].
In addition to the computation energy consumption discussed above, a significant chunk of the total energy consumption of a data center is contributed by the cooling devices, AC/DC transformation devices, etc. We consider the total computation energy as the total energy of the system and use this throughout the thesis.
1.7.2 Impact of high power consumption
High energy consumption in the servers and data centers has many demerits, and this can be categorized in three directions: (i) economic, (ii) performance, (iii) environmental. The energy consumption of a data center is almost equal to that of 25,000 households and it is around 2.2% of total electrical usages [48]. In addition, the energy consumption cost of a data center is increasing significantly. This yields a high operating electricity cost. Furthermore, high energy consumption imposes a higher cooling cost. Data center owners need to spend a significant portion of their budget for powering and cooling their servers. For instance, it is reported that Amazon spends almost 50% of their management budget for powering and cooling the data centers [57]. In [58], Koomey presented that the total power
drawn by the servers is increasing every year and if the rate continues, then the server’s energy consumption will exceed its hardware cost [59].
The second factor is the performance. High energy consumption increases the system temperature and it is reported in [60] that with an increase in every 10◦C in temperature, the failure rate of an electronic device doubles. The last factor is the adverse impact on the environment due to the emission of CO2. In addition to the monetary and performance issues, high energy consumption increases the CO2 emissions and contributes to the global warming [61]. In the year 2007, it is reported that the ICT (Information and Communication Technology) industry contributed about 2% of the global CO2emissions [62] and it is expected to increase by 180% till 2020.