An Energy Aware Cost Effective Scheduling Framework for Heterogeneous Cluster System

Highlights

• An Energy Aware Cost Effective Framework (EACEF) is proposed for scheduling in heterogeneous cluster environment.

• The system model uses DVFS technique to minimize energy consumption with green SLA.

• Resource replacement module is introduced to reduce the computation cost ($).

• Model utilizes the heterogeneous characteristics of jobs and computing resources to minimize energy and cost.

• Effectiveness of EACEF is verified through simulation using real heterogeneous PEs on two schedulers (HEFT and ESHMP).

Abstract

Cloud computing is an emerging market place in which computing resources are treated as the utilities and priced for their usage. A huge competition prevails among the cloud Service Providers (SPs) to offer eco-friendly and economically viable cloud services. Heterogeneous Cluster System (HCS) forms the backbone of such computing resources and for eco-friendly services, minimal energy consumption in HCS is a challenging issue. It not only brings down the operational cost but enhance the system reliability and other environmental veneration as well. Commonly, a green Service Level Agreement (SLA) is made between cloud users and SP. Green SLA ensures that a certain percentage of $makespan$ of the schedule is sacrificed for an energy efficient schedule. This work proposes a scheduling framework to execute heterogeneous independent jobs on a HCS. It also introduces a SLA based negotiation scheme for makespan, energy and cost referred as “Makespan, Energy and Cost Negotiator (MECN)”. The proposed model utilizes all possible empty slacks of the schedule running on Processing Elements (PEs) and extended $makespan$ (due to SLA). Using Dynamic Voltage and Frequency Scaling (DVFS) technique, it provides an energy efficient schedule. A resource replacement strategy is also incorporated, in the proposed framework, to generate a cost effective schedule without sacrificing its performance. The performance of the model is analyzed by its simulation which reflects its usefulness and effectiveness.

Introduction

Over the past decade, a remarkable growth has been witnessed in cloud computing for its on demand IT services. The success of cloud computing is significantly due to Information and Communication Technology (ICT) and its capability to fulfill the worldwide demand for dedicated and customized services  [1]. In offering these services the amount of energy consumed by the high-end computing systems, especially large-scale cluster system is significant. According to an estimate  [1], [2], high-end computing facilities in the cloud systems consume 2% of world’s electricity which is likely to grow four fold by 2020. High energy consuming resources e.g. datacenters, generate more heat requiring cooling devices which eventually emit intolerable carbon harming the environment  [3]. This raises very serious concerns regarding economic viability, environment and resource sustainability in the near future  [4], [5]. Therefore, scientists emphasize for better energy consumption of these High Performance Computing (HPC) data centers and consider it an important research topic  [6].

Modern computing devices (processors/cores/PEs) are well-equipped with the DVFS technology  [4], [7] which allows the computing resources to operate at multiple levels of voltage and frequency. DVFS technique is applied for minimal energy consumption by scaling down the voltage and the frequency levels of the involved PEs. This slows down the computation speed of the PEs, affecting the $\mathrm{makespan}$ of the schedule. To overcome this, a green SLA can be made between a client and SP which allows the client to compromise the performance ($\mathrm{makespan}$) a bit in lieu of an energy efficient schedule. Moreover, it has been observed that it is nontrivial to schedule the heterogeneous jobs to the heterogeneous PEs uniformly. In other words, some empty slacks remain unused on most of the PEs due to heterogeneous nature of jobs and PEs  [6], [8]. A possibility arises to minimize the energy consumption on HCSs using these slacks. For this, DVFS technique can be applied to scale down the supply voltage and frequency levels of the PEs for the available slack periods.

A good number of models have been proposed by the researchers that utilizes the DVFS technique to reduce the energy consumption of the HPC systems  [9], [10], [11], [12], [13]. In general, to offer an energy efficient schedule at a lower cost is strategically desired by the SP to attract more number of customers. Most of the recent works, in the domain of energy and cost saving, consider variable energy price, peak energy charge and energy consumed by the cooling systems to reduce electricity cost for the HPC cloud providers. The fundamental idea, of these energy minimization models, is to utilize the slack time in order to scale down the supply voltage and operating frequency of the computing resources [6], [8], [9], [10], [11], [12], [13]. Further, most of the research to minimize energy prices is separately done by considering various factors e.g. peak energy charges, energy consumed by the cooling system etc.  [1].

Unlike others, the proposed work utilizes the heterogeneity characteristics of the available PEs to provide an energy efficient as well as cost effective schedule. As obvious, PEs operating on higher voltage and frequency levels (provide fast computation) are offered at higher prices ($ per time unit). On the contrary, PEs operating on lower voltage and frequency levels (provide slow computation) are offered at low prices. In general, client initially demands PEs operating at the high voltage and frequency for the faster computation. Later, with green SLA, client has to sacrifice some performance, and faster PEs slow down by scaling down their voltage and frequency levels using DVFS technique. Also, it is observed that a client has to pay more for the high speed PEs which are running on low speed (poorly utilized). An opportunity exists to replace the high speed and priced ($) PEs with other available low speed and priced ($) PEs in the cluster. In order to replace PEs, it is ensured that PEs operate at appropriate voltage and frequency levels while satisfying the SLA and offer comparatively low prices.

The objective, in this work, is to offer a combined Energy Aware and Cost Effective Framework (EACEF) in presence of SLA without compromising the Quality of Service (QoS). In this framework, clients negotiates with the SP for the QoS by requesting for minimal energy consumption on a given $\mathrm{makespan}$ extension rate. Further, EACEF utilizes the heterogeneity characteristics of PEs and jobs to explore the available slacks  [6]. During slack periods, DVFS technique is used to minimize the energy consumption. A resource replacement module, an important component of the model, offers an economical schedule as per resource availability. A simulation study of EACEF is conducted on a heterogeneous cluster using three different real PEs  [8]. A workload generator is used to generate heterogeneous workload of some common mathematical and computational problems  [14], [15]. The performance of the proposed model is evaluated and is compared with two well-known heterogeneous schedulers i.e., best effort scheduler (HEFT)  [16] and ESHMP  [17].

The significant contributions, of the proposed model, are listed as below.

• It is a green SLA based novel model which minimizes energy and cost to generate an economic and effective schedule for heterogeneous cluster system. It utilizes the heterogeneity characteristics of both (PEs and jobs) to minimize energy consumption.

• It also utilizes empty slacks of the PEs created due to heterogeneity characteristics of the PEs, jobs and other slacks created due to green SLA.

• A resource replacement mechanism is also introduced to replace the higher costing PEs with lower cost PEs (as per availability). This offers a cost effective schedule without compromising the QoS.

• A simulation study is done to study the effectiveness of EACEF over two other heterogeneous schedulers i.e., HEFT and ESHMP. The comparative results show its effectiveness.

Rest of this paper is organized as follows. In Section  2, some related work is presented. The system model is introduced in Section  3 whereas problem formulation is given in Section  4. In Section  5, the proposed model is discussed. Simulation experiments for the comparative analysis is done in Section  6. Finally, the work is concluded in Section  7.