Introduction

Cloud computing is an innovative distributed computing paradigm that is widely accepted by public and private organizations. The main objective of providers is to obtain maximum profits and guarantee QoS requirements of customers. One of the main concerns is energy expenditures. Intelligent job allocation strategies can be used to improve energy efficiency.

Inefficient resource management has a direct negative effect on performance and cost. In the shared environments, it is often difficult to optimize energy consumption of physical resources and virtual machines (VMs) with different type of tasks (CPU intensive, disk I/O intensive, memory intensive, network I/O intensive, etc.). Detailed energy management at granular levels should be used to optimize resource usage and improve profitability [1].

In this paper, we present an energy optimization model in Cloud computing that takes into account different types of applications. We propose a heterogeneous job consolidation algorithm for power aware scheduling to optimize energy consumption. We evaluate power efficiency of our strategy and compare it with the best in the literature under different scenarios.

The paper is structured as follow. The next section reviews related work on the energy optimization. Section 3 presents the problem definition, while the proposed scheduling algorithms are described in Section 4. Section 5 concludes the paper by presenting main contribution and future work.

Related work

Reducing energy consumption in Cloud computing has emerged as one the main research issues both in industry and academia. This is due to the fact that the energy required by the datacenters for its operation, power supply, and cooling, contribute significantly to the total operational costs.

In this section, we discuss power aware resource allocation algorithms presented in the literature.

EMVM-Energy-aware resource allocation heuristics for efficient management [2]. The authors define an architectural framework and principles for energy-efficient Cloud computing. They present resource provisioning and allocation algorithms utilizing the dynamic consolidation of VMs for energy-efficient management of Cloud computing environments. The approach is validated by conducting a performance evaluation study using the CloudSim toolkit. It is shown that the approach leads to a substantial reduction of energy consumption in Cloud data centers in comparison to static resource allocation techniques.

Presented power consumption model is the following.

where is the maximum power consumed when the server is fully utilized; is the fraction of power consumed by the idle server (i.e. 70%); and is the CPU utilization.

The total energy consumption is defined as an integral of the power consumption function over a given period of time When VMs do not use all provided resources, they can be logically resized and consolidated to the minimum number of physical nodes. While idle nodes can be switched to the sleep mode to eliminate the idle power consumption and reduce the total energy consumption by the data center.

Fig. 1 shows the percentage of energy consumption due to CPU utilization used in this work. HSFL-Hybrid shuffled frog leaping algorithm [3]. The authors propose a data center resource management scheme. It can not only guarantee user quality of service (QoS) specified by SLAs, but also achieve maximum energy saving and green computing goals. Consolidation of resources is achieved by VM migrations technology. Low utilized and idle hosts are switched to power saving mode to achieve energy saving while ensuring that SLAs are adhered to.

Host energy consumption exhibits an almost linear proportion to CPU energy consumption. Moreover, the energy consumption of an idle host accounts for 70% of full-load operation energy consumption. The energy consumed by VM migrations also requires consideration. Energy consumption within a given unit time is defined as follows is the energy consumption when host is in full load.

is the average utilization rate of the host processor within unit time, is the collection of VM migrations within the unit time window, and is the migration time of VM . The percentage of energy consumption due to CPU utilization is similar to Fig. 1.

AETC-Algorithm of energy-aware task consolidation [4]. The authors propose a technique of energy-aware task consolidation (ETC) to minimize energy consumption. ETC restricts CPU use below a specified peak threshold by consolidating tasks amongst virtual clusters. In addition, the energy cost model considers network latency, when a task migrates to another virtual cluster. They define a default CPU utilization threshold of 70% to demonstrate task consolidation management amongst virtual clusters. Although the idle state of virtual machines and network transmission are assumed to be a constant ratio of basic energy consumption unit in his study. The simulation results show that ETC can significantly reduce power consumption when managing task consolidation for Cloud systems. ETC is designed to work in a data center for VMs that reside on the same rack or on racks where network bandwidth is relatively constant. The model assumes energy consumption in the idle state. An additional energy β is required for executing tasks when CPU utilization is increased.

The energy consumption of a virtual machine is defined as follows. The total energy consumption of during the time period is given the following formula: Given a virtual cluster, which consists of VMs, the energy consumption of VC during the time period tm is as follows:

Fig. 2 shows the percentage of energy consumption due to CPU utilization. CTES-Cooperative Two-Tier Energy-Aware Scheduling [5]. The authors propose a cooperative two-tier task scheduling approach to benefit both Cloud providers and their customers. It regulates the execution speeds of real-time tasks in a way that a host reaches the optimum level of utilization instead of migrating its tasks to other hosts. They also propose several predictive global task scheduling policies to map arrived tasks to feasible VM, in his technique, a host is locally scheduled to reach its optimum CPU usage instead of migrating its tasks to other hosts. They divide the energy consumption of a host into two parts, static and dynamic energies. His simulation results show that the proposed task scheduling approach reduces the total energy consumption of a Cloud.

The utilization of a host, is defined as:

, where are the allocated MIPS of in time and is the maximum computing power of We suppose that an idle host changes its state to be powered off immediately. Thus, the total power of a host is defined as: is the power consumed during the idle time of a computing node. It is defined as .

is the power consumed when a host works with its maximum utilization. Utilization. is the constant ratio of the static power of a host to its maximum power which depends on the physical characteristics of a host.

Dynamic power consumption is:

If the system uses the power , the energy consumption will be where is the time in which a host works at its maximum computing power to finish a certain number of instructions. The percentage of energy consumption due to CPU utilization is similar to Fig. 1.

Therefore, the energy consumption of a host to finish its certain amount of instructions is obtained by: DVMA-A Decentralized Virtual Machine Migration Approach [6]. The authors propose a decentralized virtual machine migration approach inside the data centers for Cloud computing environments which use virtual machines to host many third-party applications. They define a system models and power models then; they present the key steps of the decentralized mechanism, including the establishment of load vectors, load information collection, VM selection, and destination determination. A two-threshold decentralized migration algorithms is implemented to further save the energy consumption as well as keeping the quality of services. Performance evaluation results of their simulation experiments illustrate that their approach can achieve better load balancing effect and less power consumption than other strategies.

An idle physical node even with 0% of utilization could still consume a plenty of power. Let α be the fraction of power consumed by an idle node compared to a full utilized node and  the current CPU utilization of the node. Then, we use the power model defined as follows to compute the power consumption of a physical nodes : , where is the power consumption of when it is fully utilized (i.e., it reaches 100% of CPU utilization). The percentage of energy consumption due to CPU utilization is similar to Fig. 1.

EDRP-Energy and Deadline Aware Resource Provisioning [7]. The authors addresses the problem of minimizing the operation cost of a Cloud system by maximizing its energy efficiency while ensuring that user deadlines as defined in Service Level Agreements are met. They take into account two types of workload models, independent batch requests and task graphs with dependencies.

The power consumption of at time includes the static power consumption and the dynamic power consumption . Both are correlated with the utilization rate of at time : . We evaluate by considering only the CPU requirements of the hosted VMs indicated in , and do not differentiate between VMs that are running tasks and idle VMs, since background CPU activities are needed even during idle periods.

is constant when , 0 otherwise. The relationship between and is much more complex. Servers have optimal utilization levels in terms of performance-per-watt, which we define as for . It is commonly accepted that for modern servers , and the increase in power consumption beyond this operating point is more drastic than when . Even for identical utilization levels, the energy efficiency of different servers may vary. This is captured by the coefficients and , representing the power consumption increase of when and respectively.

is then calculated as:

We would like to point out that the exact formulations of do not undermine the analysis, since its increment is faster when than when .

Suppose the upper bound of the maximum schedule length of all applications is . The total energy consumption (COSP) is the sum of the power consumption across all servers throughout the operation timeline:

In Fig. 3, we show the nonlinear percentage of energy consumption due to CPU utilization used in this work. [8]. The authors propose a simulation-driven methodology with an energy model based on polynomial regression with Lasso to predict energy consumption to verify its performance, and a resource scheduling algorithm BFDP shifting its optimization goal from resource consolidation to power consumption to improve the energy efficiency without degrading the QoS of the system and they consider four type of jobs, CPU-intensive, Memory-intensive, Networkintensive and I/O-intensive. The authors introduced the mechanism of utilization thresholds in BFDP to alleviate the over-consolidation issue in the Best-Fit strategy. Their results showed that are effective because BFDP creates less SLA violations than the BFDR in light workloads.

BFDP-Best Fit Decreasing Power

The authors uses a nonlinear energy model: where is the kernel function of expression , is the CPU and memory utility. is the parameter of the kernel function to be determined through the model training process. is a constant. Fig. 4 presents a relationship between CPU and memory utilization and full-system power. Суперкомпьютерные дни в России 2015 // Russian Supercomputing Days 2015 // RussianSCDays.org they use Xen as the Virtual Machine Monitor to make experiments. Under different resource allocation configurations, they evaluate power efficiency up to 2%12%, compared to the default deployment. Finally they conclude that if the CPU-Intensive application is allocated twice as much CPU compared to I/O-Intensive application, there are an improvement in the power efficiency.

Table 1 shows the summary of the algorithm domains, the main characteristics of described algorithms, and the criteria used to evaluate quality of the algorithms.

• • • • • • • • • • • [2] HSFL • • • • • • • • • [3] AETC • • • • • • • • • [4] CTES • • • • • • • • • • [5] DVMA • • • • • • • [6] EDRP • • • • • • • • • • • [7] BFDP • • • • • • • • • • [8] PAHD • • • • • • [9]

Problem definition

We assume that servers of the data center are identical, and described by tuples { , , , , }, where is a measure of instruction execution speed (MIPS), is the amount of memory (MB), is the available bandwidth (Mbps), and is energy efficiency (MIPS per watt). We also assume that data centers have enough resources to execute any job.

The main objective of the proposed strategies is to minimize the total power consumption of running workloads providing QoS guarantees.

Job model

We consider independent jobs . The job is described by a tuple , where is the released time, is a processing time. The release time of a job is not available before the job is submitted.

is the SLA from a set offered by the provider [19,10,20]. Each SLA represents a SL guarantee, and is denoted by the slack factor . is the deadline of the job and is calculated at the release of the job as . Finally characterizes a job as CPU intensive, disk I/O intensive, memory intensive, network I/O intensive, etc.

Energy model

We present a nonlinear model of the power consumption by considering types of applications. Fig. 5 shows examples of the normalized power consumption of jobs of type A and type B vs CPU utilization (%). Characteristics of CPU intensive, disk I/O intensive, memory intensive, network I/O intensive, etc. applications influence on power consumption differently due to corresponding hardware characteristics.

Moreover, an allocation of two different applications to the same server could cause reduced power consumption, less than the sum of their individual power consumptions. It also has an impact on performance enhancement avoiding creation of a bottleneck and resource contention (either in CPU, disk or network) that may result in additional degradation of the system performance and increased energy consumption.

We propose a hybrid model that takes into account power consumption of individual jobs and their combinations. Due to diversity of applications and their combinations, we propose to consider aggregated utilization of each type of applications (total utilization that contributes each job type or concentration). The power consumption of the processor at time consists of two parts: an idle power consumption when the processor is turned on, but not used , and power consumption when the processor is in use :

,

where , if the processor is on at time , and otherwise.

is the utilization at time . is a coefficient proposed in [15] to fit non-linear power profiles. ,

where is the maximum power consumption when the processor is fully utilized. is the coefficient that represents the increment of power consumption when a processor runs different types of applications. The concentration of type A vs type B at the time is defined as .

(

Fig. 6 shows the proportion of power consumption, when the processor runs different jobs A and B. 0 The total power consumed by the system is the integral of power consumed during operation:

, with

We define . Following [11], with the power consumption of a processor Fujitsu PRIMERGY TX300 S7, we set . We set , for (all jobs are type A), and for (all jobs are type B). Fig. 7 shows the normalized power consumption when the processor runs two types of applications.

Scheduling algorithms

In this section, we describe our scheduling approach and proposed energy-aware scheduling methods.

Scheduling approach

We address a two-level scheduling approach [12,13,14,15]. At the upper level, the system verifies whether a job can be accepted or not using a Greedy acceptance policy. If the job is accepted then the system selects a machine from the set of admissible machines to execute the job on the lower level. The greedy higher-level acceptance policy is based on the preemptive Earliest Due Date (EDD) algorithm, which gives priority to jobs according to their deadlines. When a job arrives to the system, in order to determine whether to accept or reject it, the system searches for the set of machines capable of executing the job before its deadline assuring that no jobs in the machine will miss their deadlines. If the set of available machines is not empty job is accepted otherwise it is rejected. This completes the first stage of scheduling.

Note that the preemptive EDD algorithm is well suited for our purpose as it is easy to apply and it yields an optimal solution for the 1 | prmp, , online | Lmax problem. By EDD, we verify that all already accepted jobs with a deadline greater than the deadline of the incoming job will be completed before their deadline.

Allocation strategies

The machine for job allocation can be determined by taking into account different criteria. In this work, we study ten allocation strategies Rand, FFit, RR, ML, MTe, Me, Mu, Mau, Mujt, and Mc. (see Table 2). They are characterized by the type and the amount of information used for allocation decision.

We categorize the proposed methods in three groups: (1) knowledge-free, with no information about applications and resources [16,17,18]; (2) energy-aware, with power consumption information; and (3) utilization-aware with utilization of machines information. allocates job to the machine with minimum concentration of jobs of the same type at time

Conclusions

In this paper, we consider the problem of energy optimization in Cloud computing from the perspective of the Cloud service provider. We formulate and discuss the energy model and energyaware resource allocation problem that provide energy-efficiency and QoS guarantees simultaneously by heterogeneous job consolidation taking into account types of applications. A generic Cloud computing environment has to process multiple applications for multiple users, which create mixed workloads of different types with different energy consumption.

We consider energy characteristics of applications such as CPU intensive, disk I/O intensive, memory intensive, network I/O intensive, etc. and their influence on power consumption due to the nature of used hardware. We discuss how an awareness of the job type could help to improve energy consumption. Intelligent job allocation has an impact on performance enhancement avoiding creation of bottlenecks and resource contentions either in CPU, disk or network, and on decreasing total energy consumption.

We propose a hybrid model that take into account the power consumption of individual jobs and their combination. We propose using aggregated utilization of applications, and their concentration for job allocation.

However, further study for energy consumption of multiple job types and their concentration is required to assess the actual efficiency and effectiveness of the proposed method. This will be the subject of future work for better understanding of the resource contentions and its impact on the energy consumption, QoS and multi-objective optimization in clouds.