Thermal aware overall energy minimization scheduling for hard real-time systems☆
Introduction
The human's appetite for high-performance computing systems has driven the semiconductor technology into the deep sub-micron (DSM) era. The continued shrinking of the transistor size, together with increasingly complicated circuit architectures, have resulted in a significant increase of the power density. It is predicted that in a matter of a few years, the number of transistors that can be integrated into one 300 mm2 die will reach 100 billion and its power consumption can be as high as 300 W [1]. One of the immediate consequences is the energy consumption issue which has posed tremendous challenge to digital system designers of both embedded and server platforms. For battery-driven embedded systems, given the fact that the growing of battery technology is far lagging behind the development of the semiconductor technology, unless more advanced low-power or power-aware design techniques are applied, the applicability of such devices will be severely limited. For power-rich server systems, on the other hand, the high energy consumption directly leads to high temperature, which not only increases the packaging/cooling costs, degrades the reliability/performance of the system, but also significantly increases the leakage power consumption due to the strong leakage/temperature dependency in the DSM domain. Therefore, power/energy minimization techniques are urgently demanded at every design abstraction levels.
Early research efforts are mainly focusing on how to reduce the dynamic energy consumption of a system. Dynamic voltage scaling (DVS) is well known to be one of the most effective dynamic energy reduction techniques. Due to the convex relationship between the dynamic power consumption and the processor speed level, a quadratic dynamic power reduction can be achieved at the expense of a linear performance reduction. However, as the leakage becomes more and more prominent in the DSM domain, the leakage energy component as well as its interdependence with the temperature have to be properly taken into consideration. As evidenced in [2], the leakage power will increase by 38% when chip temperature raising from 65 °C to 110 °C. In fact, thermal/temperature constraint is becoming an increasingly critical issue in the computing system deign [3]. Due to the leakage/temperature dependency, the leakage energy consumption is as important as the dynamic energy consumption. Therefore, an energy optimization design technique can become ineffective or inefficient if the interdependency between leakage and temperature is not properly addressed.
In this paper, we are interested in studying the problem of energy minimization for hard real-time scheduling on single-core systems by considering the dependency between leakage and temperature. The complexity of the problem lies in the fact that leakage energy consumption depends on both supply voltage level and temperature. For instance, two identical speed schedules may result in different energy consumptions simply because their initial temperatures are different. Therefore, how to accurately and efficiently estimate the energy consumption of various candidate speed schedules is the key to solve this problem.
We summarize the major contributions of this work below:
- •
First, we develop an effective closed-form energy estimation method with the leakage/temperature dependency taken into consideration.
- •
Secondly, based on the proposed energy estimation method, we further develop two scheduling algorithms to minimize the overall energy consumption. The first algorithm is an off-line algorithm, targeting at a real-time system consisting of a set of periodic tasks with the same period and deadline. The second algorithm is an on-line algorithm, which is intended for more general real-time systems consisting of multiple sporadic tasks.
- •
Finally, we conduct extensive experiments to evaluate the performance of our proposed energy estimation method as well as the online/offline energy efficient scheduling algorithms. Our experimental results show that the proposed energy estimation method can achieve a speedup up to two orders of magnitude compared with the existing approach while maintaining high accuracy. Our experimental results also show that the proposed scheduling algorithms consistently outperform the existing approaches in terms of energy reduction.
The rest of this paper is organized as follows. Section 2 discusses the related work. Section 3 introduces the system models used in this paper and Section 4 provides a motivational example. In Section 5 we show how to derive the proposed energy estimation equations, based on which we present our overall energy minimization scheduling algorithms in Section 6. Experimental results are discussed in Section 7. Section 8 concludes this paper.
Section snippets
Related work
The power/energy minimization problem has been researched extensively for a couple of decades [4], [5], [6], [7], [8], [9], [10]. Early research efforts are mainly focusing on reducing the dynamic power/energy, i.e. the predominate component of the overall power/energy consumption. By taking advantage of the convex relationship between the dynamic power and supply voltage, many technique (e.g. [4], [5]) strive to lower down the processor supply voltage and working frequency. As the leakage
System models
In this paper, we choose the processor model as single-core systems. For the real-time applications, we first consider a set of hard periodic tasks with the same period, or equivalently, one single consolidated task with the period equal to its deadline. We further extend our research to a more general real-time application model, in which task set (Γ) consists of n sporadic tasks, denoted as Γ = {τ1, τ2, …, τn} and scheduled according to the EDF policy. For each task τi ∈ Γ, we have τi ≡ {ai, ci, di
Motivational examples
As leakage varies with temperature, one key challenge in our research is to calculate the overall energy consumption of a speed schedule accurately and efficiently.
Let us consider a speed schedule shown in Fig. 1(a). The speed schedule, S ≡ [A, B, C, D], consists of four intervals with each characterized by a speed level (s1 to s4) and a duration ((t1 − t0) … (t4 − t3)).
Since the leakage energy used to take only a small portion of the overall energy consumption, some early works [4], [5], [6], [7], [30]
Energy estimation equation
From the motivational example above, we can see that it is critical to develop a method that can rapidly and accurately estimate the energy consumption of a given schedule. As one of the major components of overall energy consumption, the leakage energy as well as its strong dependency with temperature have made the energy estimation a non-trivial task. In this section, we first discuss how to calculate the energy consumption for a single speed interval. Then, we discuss how to calculate the
Our energy minimization scheduling algorithms
In the previous section, we proposed a simple yet effective energy estimation method to calculate the overall energy consumption for a given schedule. Then, by formulating the differences between the ending and the starting temperature of each consecutive copy of a periodic schedule as geometric series, we further proposed an efficient method to estimate the overall energy consumption of a given periodic schedule.
In this section, based on the proposed energy estimation methods and the so-called
Experimental results
In this section, we use experiments to evaluate our methods. We first examine the accuracy and efficiency of the proposed energy estimation equations. Then, the performance of the proposed energy minimization scheduling techniques will be tested and discussed.
Conclusions
In this paper, we study the problem on how to reduce the overall energy consumption of a hard real-time system by applying real-time scheduling techniques with the leakage/temperature dependency taken into consideration. We first develop an energy estimation method that can be used to accurately and efficiently calculate the overall energy consumption of a candidate schedule. Then, based on the proposed energy equation, two scheduling approaches are proposed. The first one is an off-line
References (33)
- et al.
An overview and classification of thermal-aware scheduling techniques for multi-core processing systems
Thousand core chips: a technology perspective
- et al.
Temperature and supply voltage aware performance and power modeling at microarchitecture level
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
(2005) - et al.
On-line dynamic voltage scaling for hard real-time systems using the EDF algorithm
- et al.
A scheduling model for reduced CPU energy
- et al.
Leakage-aware energy-efficient scheduling of real-time tasks in multiprocessor systems
- et al.
Dynamic slack reclamation with procrastination scheduling in real-time embedded systems
DAC
(2005) - et al.
Speed scaling to manage energy and temperature
Journal of the ACM
(2007) - et al.
Guaranteed scheduling for repetitive hard real-time tasks under the maximal temperature constraint
- et al.
Energy-efficient variable-flow liquid cooling in 3d stacked architectures
Leakage temperature dependency modeling in system level analysis
Temperature-aware idle time distribution for energy optimization with dynamic voltage scaling
On-line thermal aware dynamic voltage scaling for energy optimization with frequency/temperature dependency consideration
Leakage aware energy minimization for real-time systems under the maximum temperature constraint
Energy-efficient real-time task scheduling with temperature-dependent leakage
Feasibility analysis for temperature-constraint hard real-time periodic tasks
IEEE Transaction on Industrial Informatics
Cited by (3)
Scheduling-based power capping in high performance computing systems
2018, Sustainable Computing: Informatics and SystemsCitation Excerpt :We therefore need a mechanism to estimate the future power consumption of a job using the information available at schedule-time. The importance of such predictions was underlined by several works [38–40]. Furthermore, a greater prediction accuracy is related to a better performance of a power capped dispatcher (in terms of higher machine utilization and greater energy savings) [41].
Optimized Cooperative and Random Schedulings Packet Transmissions and Comparison of Their Parameters
2018, Wireless Personal CommunicationsAn energy-aware dynamic scheduling algorithm for hard real-time systems
2014, Proceedings - 2014 3rd Mediterranean Conference on Embedded Computing, MECO 2014 - Including ECyPS 2014
- ☆
This work is supported in part by NSF under projects CNS-0969013, CNS-0917021 and CNS-1018108.