Energy efficient duplication-based scheduling for precedence constrained tasks on heterogeneous computing cluster

The continually increasing energy consumption represents a critical issue in modern heterogeneous computing systems. With the aid of dynamic voltage frequency scaling (DVFS), task scheduling is considered an effective software-based technique for reducing the total energy consumption and minimizing the overall schedule length (makespan). A natural solution is to reclaim the slack time in a given time-efficient schedule, which is also referred to as a "two-pass" method or a "rescheduling" method. A number of studies have focused on slack reclamation to achieve energy reductions through heuristics; although, these methods offer suboptimal solutions. In this article, the rescheduling optimization problem is formulated as a linear program for minimizing an energy objective function subject to precedence and deadline constraints implied in the given schedule. Two types of decision variables, ie, frequency duty factors and task intervals, are defined to set up the linear model. Consequently, an optimal solution to the problem can be provided in a straightforward manner by a linear programming solver, which suggests that such a rescheduling problem belongs to the P (polynomial time) class. The experimental results show the effectiveness of the proposed approach and demonstrate that the performance is superior to that of other competitive algorithms in terms of both energy saving and runtime efficiency. KEYWORDS dynamic voltage frequency scaling (DVFS), energy efficiency, heterogeneous computing system, linear programming, slack reclamation, task scheduling 1 INTRODUCTION A heterogeneous computing system (HCS) is a computing platform with diverse sets of heterogeneous computing resources connected by a high-speed network for processing parallel applications. 1 In recent decades, HCSs have been widely used in both scientific and commercial fields, such as supercomputing, cloud computing, and big data. For example, the world's fastest supercomputer "Sunway TaihuLight" has 40,960 computer nodes, each with 4 management processing elements (MPEs) and 256 computing processing elements (CPEs). Although it provides the most powerful computing capability, "Sunway TaihuLight" also consumes a tremendous amount of electrical energy, estimated at approximately 15,371 kW. 2,3At an approximate price of $0.1/kWh, its energy cost is approximately $13.46 million per year, which is far beyond the acceptable cost for many HCS operators. Furthermore, the energy cost of powering a typical data center doubles every five years. 4 In certain cases, the power cost may exceed the hardware purchase costs. 5 Besides, high energy consumption creates a number of environmental problems. 6,7 For example, in 2014, global data centers consumed up to 3% of the world's electricity production while causing 200 million metric tons of carbon emissions, thus accounting for approximately 2% of the world's greenhouse gas emissions. 8 On the other hand, a large portion of servers inside these data centers have relatively low average utilization efficiency. Observa...

Section: Of 17mentioning

confidence: 99%

Energy‐efficient task scheduling on heterogeneous computing systems by linear programming

Zhang

Wang

Tang

et al. 2018

“…Mostly, the research works in multiprocessor scheduling explored DVFS technique extensively for reducing dynamic energy consumption. [10][11][12][13]20 Many list-based heuristics exploited task scheduling with DVFS to reduce energy consumption. 10,12,13 However, listbased energy-aware algorithms were quite effective in reducing dynamic energy consumption but overlooked communication energy of interconnects.…”

Section: Related Workmentioning

confidence: 99%

“…Judicious task scheduling is one of the vital research issues in parallel processing systems and has been extensively studied by the research community. [10][11][12][13] Apart from computation energy, network's communication energy is also important particularly for communication intensive graphs 14,15 and thereby needs to be handled explicitly. These systems compose of autonomous dissimilar processors interconnected by high-speed networks.…”

Section: Introductionmentioning

confidence: 99%

“…In former technique, the basic idea is to put the system components (processors, hard drives, network cards, etc) to one of the low-power states for underused time (scheduling slots/idle time slots) to reduce static power consumption, 4,8,9 while, in the latter, processor's operating voltage V and frequency f are scaled down for scheduling slots leading to reduce computation (dynamic) energy consumption. [10][11][12][13] Apart from computation energy, network's communication energy is also important particularly for communication intensive graphs 14,15 and thereby needs to be handled explicitly. For example, 15 projected energy trend of Internet resources feed to routers, switches, and servers was about 868 billion KWh in 2008 and is predicted to increase 5 times by 2025.…”

Section: Introductionmentioning

confidence: 99%

“…Dynamic power management (DPM) and dynamic voltage and frequency scaling (DVFS) are 2 widely adopted and efficient techniques to deal with optimum energy efficiency. In former technique, the basic idea is to put the system components (processors, hard drives, network cards, etc) to one of the low‐power states for underused time (scheduling slots/idle time slots) to reduce static power consumption, while, in the latter, processor's operating voltage V and frequency f are scaled down for scheduling slots leading to reduce computation (dynamic) energy consumption …”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Duplication‐controlled static energy‐efficient scheduling on multiprocessor computing system

Kaur

Bansal

2017

Self Cite

Energy-efficient scheduling is a step towards meeting green computing requirements. The work in this direction mainly aims at reducing dynamic energy consumption that includes clock gating, cache subbanking, and dynamic voltage and frequency scaling of underlying processors. However, the emergence of fast and compact transistor sizes has exponentially added onto the processor static power consumption, which has not been paid much attention. This article proposes a duplication-controlled static energy-efficient scheduling (C-SEED) algorithm for scheduling precedence constrained applications on parallel computing systems. The C-SEED algorithm couples adaptive threshold-based duplication with system level dynamic power management technique to achieve its objectives. Dynamic power management works by selectively putting the energy-consuming resources to efficient low-power states for idle times to reduce energy consumption. Efficacy of the proposed algorithm is analyzed and compared against other relevant works on the basis of makespan and total energy (dynamic + static + communication)consumption. The extensive simulation results carried over large set of random and regular task graphs show that the proposed C-SEED algorithm has potential to reduce energy consumption as well as makespan.KEYWORDS duplication-based scheduling, dynamic power management, list-based scheduling, multiprocessor scheduling, static power consumption

A variable neighborhood search algorithm for energy conscious task scheduling in heterogeneous computing systems

Zhang

Tong

et al. 2021

Energy efficiency in heterogeneous computing systems has attracted increasing interests due to its economic and environmental impacts during recent decades. Based on power-aware hardware techniques, such as dynamic voltage frequency scaling, efforts have been made through task scheduling to reduce the total energy consumption for executing a parallel application while maintaining its time efficiency. In this case, energy conscious task scheduling refers to a bi-objective optimization that aims to minimize the overall completion time (makespan) and the total energy consumption, simultaneously. Existing energy conscious scheduling algorithms conduct energy optimization by means of slack reclamation or a trade-off function. However, the performance of slack reclamation has been proved to be upper-bounded and methods relying on trade-off functions cannot guarantee bi-objective optimization. In this article, an energy conscious task scheduling algorithm is proposed to tackle the above issues based on the framework of variable neighborhood search. Two neighborhood structures are designed to reduce makespan and the total energy consumption, respectively. Furthermore, a pruning technique is incorporated into the algorithm to accelerate the searching process. Extensive experimental results on both randomly generated and real-world applications demonstrate that the proposed algorithm improves the time-efficient schedules on average by 22.4% for the energy consumption and 1.2% for the makespan. K E Y W O R D Sdynamic voltage frequency scaling (DVFS), energy efficiency, heterogeneous computing system, task scheduling, variable neighborhood search (VNS)