2.2 Optimization Objectives of Resource Allocation
2.2.2 Energy Consumption
2.2. OPTIMIZATION OBJECTIVES OF RESOURCE ALLOCATION
total processing time oftj, the time periodtj is running alone and thattj running in parallel with one or more tasks, respectively.
There is another dynamic power management technique called Dynamic Voltage and Frequency Scaling (DVFS) is used to control energy consumption. The power con- sumption of a processor depends on the operating frequency of the CPU. Lowering the operating frequency of the CPU may lead to power savings and potential energy savings but may impact application performance, which means that the same applica- tion may need more time to complete execution [193]. So to increase energy efficiency while meeting the SLA constraints, the scheduling algorithm must determine the op- timal frequency for each application. They used the following energy model for their work.
ET(s) =EdynT (s) +EstaticT (s) (2.3) where EdynT (s) is the dynamic power consumption and EstaticT (s) is static energy con- sumption of the system.
Lee et al. [142] have proposed makespan-conservative energy reduction along with simple energy conscious scheduling to find a trade-off between the makespan time and energy consumption, where they reduced both makespan time and energy con- sumption of precedence constraint graph on heterogeneous multiprocessor systems supporting DVFS technique.
Ferdaus et. al. [96] proposed an ant colony optimization approach to achieve workload balancing in a cloud environment. They use a linear resource utilization model along with the following energy model of each application. The energy model was defined as follows.
E(p) =
Eidle+ (Ef ull −Eidle)×UpCP U, if UpCP U >0
0, Otherwise
where Ef ull and Eidle are the average energy drawn when a PM is fully utilized (i.e.
100% CPU busy) and idle, respectively, and UpCP U represents the CPU utilization.
Another way to improve energy efficiency is to switch-on the server when resource needs increase and switch-off the servers when not in use. However, deploying the optimal number of servers at run time by understanding the workload dynamics is a challenging task. Zhang et al. [257] proposed a framework that dynamically
2.2. OPTIMIZATION OBJECTIVES OF RESOURCE ALLOCATION Table 2.2: Representative works based on energy efficiency
Ref. Main objective Technique used Evaluation [128] Improve the energy
consumption and CO2 emissions
Multi-objective genetic algorithm (MO-GA)
Experimentation using Feitelson’s PWA Parallel Workloads [126] To improve resource
utilization, perfor- mance improvement and energy effi- ciency
Artificial Bee Colony Experimentation using simulation tool, CloudSim with random workload [208] To optimize execu-
tion time, cost and energy consumption satisfying the QoS requirements
Compromised cost-time based (CCTB) scheduling policy, time based (TB) scheduling policy, cost based (CB) scheduling pol- icy and bargaining based (BB) scheduling policy
Experimentation using CloudSim with randomly generated work- load
[196] To optimize lifetime reliability of appli- cation and energy consumption with guarantees of QoS constraint
Heuristic of Reliability and Energy Efficient Workflow Scheduling (REEWS)
Simulation us- ing CloudSim results obtained by using ran- domly generated task graphs [213] Reduce physical
machine energy consumption and communication cost
Genetic algorithm Simulation us- ing synthetic dataset [103] Reduce resource
wastage and energy consumption
Use multi-objective algo- rithm
Numerical simu- lations
determines the number of machines required and adjust the resource provisioning un- derstanding the trade-off between energy consumption and scheduling delay. Their proposed framework considers the heterogeneity of workload where tasks are clus- tered based on their requirements and resource needs. They subsequently adjust the placement to heterogeneous PMs taking into account the reconfiguration cost.
In a cloud environment, a significant amount of energy is consumed by cooling sys- tems. When the load of the system is high that causes thermal hotspots and hence needs more energy for cooling the system. So the challenge for the researchers is to distribute the load uniformly among the servers to avoid thermal hotspots. Sand-
Table 2.3: Representative works based on cost optimization objective Ref. Main objective Technique used Evaluation [176] To improve upon
total cost, time com- plexity, and schedule length
A novel hybrid algorithm, called CR-AC, combining both the chemical reaction optimization (CRO) and ant colony optimization (ACO) algorithms to solve the workflow-scheduling
CloudSim toolkit and evaluated by using real ap- plications and Amazon EC2 pricing model [121] Proposed model to
reduce the makespan, cost and deadline violation rate
Hybrid algorithm com- bined two optimization algorithms namely called as Cuckoo Search (CS) and Particle Swarm Optimiza- tion (PSO)
Use of Cloudsim toolkit with randomly generated workload [234] To optimize
makespan, cost and CPU time
Metaheuristic based scheduling algorithms including genetic algorithm (GA), ant colony optimi- sation (ACO), and particle swarm optimisation (PSO) are adapted
Implemented in SwinDeW-C cloud work- flow system to demonstrate the perfor- mance [72] To optimize cost and
makespan
Uses dynamic objective GA (DOGA) with adap- tive ability to the search environment
Experiment with workflows
[31] Improve the makespan time, throughput, availabil- ity and cost
Load balancing resource clustering (LB-RC) algo- rithm using BAT optimiza- tion technique
Simulation using synthetic dataset
piper [233] uses resource usage data of VMs through profiling and use that data to detect the hotspots. Hotspots are reactively mitigated using VM resizing or migration from overloaded servers to less loaded servers. Some representative works concerning energy consumption is presented in Table 2.2. The detail survey on energy-efficient scheduling are presented in these papers [34], [127], [164].