Energy-Efficient Scheduling Based on Task Migration Policy Using DPM for Homogeneous MPSoCs

1  Introduction

The dissipation of energy becomes the captious design constraint of system-on-chip (SoCs) over the past decade as it limits the performance, reliability and battery life. Due to advancements in computational embedded devices and increasing multi-task execution than ever before [1].

MPSoC is widely deployed in high-performance computing and application-specific embedded systems such as gaming and aerospace-based systems for real-time response. It contains high-performance ARM Cortex-A7, ARM Cortex-A15 and an energy-efficient INTEL PXA-270 processor [2]. Introduced a technique that reduces resistance and energy because lack of concentration can affect the reliability and life span of the chip as well as overall systems performance [3]. Due to the high processing of tasks temperature on the chip increases. Various mechanisms are used to reduce the thermal effects due to high heat and decrease the performance of the system because high heat causes the chip to be damaged. A central processing unit is working as the main processing unit for performing the instructions read and write operation. CPU unit is placed in all the modern embedded systems [4]. The processing unit needs to be updated with time if the processing unit has a higher processing speed it can execute and manage intense tasks efficiently at short intervals of time. Advancement in the processing unit makes our system run heavy tasks but it can have some issues e.g., Dimension, cost, energy, power utilization, performance, reliability and processing speed. Switching of the task is the major issue with the evolvement of the processor [5]. DPM mainly deals with the development of policies that analyze the run time behavior of the system to reduce the power consumption of the MPSoC system [6].

Introduced simulated annealing based (LPPWUSa) optimization strategies for reducing the energy and solving the system-level low power design problems [7]. Proposed a dynamic thermal management (DTM) based energy optimization technique for the MPSoC platform with dynamic voltage and frequency scaling (DVFS) enabled homogeneous processors. These are one of the most reliable techniques to reduce and stabilize the temperature of the multi-core system. In a multi-core system, an exponential decrease in the temperature also reduces the power utilization [8]. Introduced homogeneous multiprocessing often known as symmetric multiprocessing (SMP) is a common type of multiprocessing in computer systems, in which two or more identical processors share a single main memory for process execution [9]. Introduced dynamic thermal management (DTM) based technique that efficiently manages the thermal responses of a processing system. Many techniques are combined to manage temperature and thermal responses including DVFS, DPM and Dynamic voltage scaling (DVS). These techniques are very useful but they cause some reliability and performance issues. They are mostly used to resolve on-chip power dissipation problems [10]. Modern MPSoCs are based on CMOS chips. The switching frequency regulates how often the switches occur. Double-edge-triggered flip-flops can be utilized to minimize dynamic power for MPSOC technology. An embedded device’s power consumption is categorized as Pdynamic and Pstatic [11].

The accumulative power of a CPU is the ∑ of the Pdynamic because of switching and the Pstatic occurs due to leakage of power. The primary factor in CMOS circuits is dynamic power dissipation that occurs because of transistor switching. Pdynamic can be calculated using Eq. (1):

Pdynamic=Cload⋅Ncs⋅Vdd⋅f(1)

f=R⋅(Vdd−Vt)/⋅Vdd.2f(2)

Load capacitance is denoted as Cload, Ncs is the number of circuit transitions while supplied voltage is denoted as V which is equivalent to switching voltage. f is the clock frequency while Vt represents the threshold voltage. R is a constant and reduction in supply voltage is represented as Vdd as shown in Eq. (2). La is the total number of logic gates in the CMOS circuit. The total power of the CMOS circuit is represented in Eq. (3):

Ptotal=(CeffVdd.2f+La(Vdd.R3.)(3)

Most of the embedded computing circuits aim to give maximum performance while using minimum power. In static power, the dissipation of power occurs when the circuit is not changing states due to leakage current. The short circuit power is utilized when both positive channel metal-oxide-semiconductor (PMOS) and negative channel metal-oxide-semiconductor (NMOS) have switched ON for a short period unless the path between supply voltage is directed with the ground [12].

2  Literature Review

A decrease in the chip size of a multi-core processor certainly increases the number of transistors on a chip more rapidly than before. The chip can require more energy due to which a gradual increase in power density is observed that affects the reliability of a multi-core processor [13]. Scheduling and task allocation techniques are facing issues during the migration of tasks to balance and manage power on multiprocessor systems. Tab. 1 represents the comparison of various online and offline DPM-based scheduling techniques [14].

Fig. 1 shows that higher power consumption can increases higher temperature and higher resistance eventually causing lower possible speed and hotspots that cause the permanent failure of the device [15].

Figure 1: High utilization effects on system on chip (SOC)

Introduced a technique that increases the life span of the chip by reducing the consumption of power in multiprocessors-based systems. DPM-based technique affects the performance of the processor by suddenly switching the system from idle to running state. An efficient dynamic power management policy keeps both the reliability and performance while considering power degradation within allowed limits. Designs of these policies are considered to be an active and burning research topic in the field of MPSoC while designing embedded systems few challenges occur such as size, cost, power consumption and reliability [16].

2.1 Dynamic Power Management

DPM technique selectively turns off all those components that are not in use. DPM is used in several portable systems but its applications are not yet explored because of the complexity of interfacing heterogeneous components. The fundamental problem in the implementation of DPM techniques is the non-uniform workload during the execution of the task. To solve this problem DPM uses a predictive algorithm that predicts the future workload by using different predictive models. It covers several system-level DPM approaches to save energy. DTM technique is used to find an optimal solution to avoid peak temperature which causes hotspots on chips [17]. Energy is not directly affected by temperature but when temperature increases from the threshold value some cooling mechanism is required to reduce the temperature of the system. Cooling mechanisms consume energy to reduce the temperature [18]. In a single-core processor if we increase the frequency by 50% that roughly increases the power consumption two times while in dual-core systems power consumption increases by 30% if we increase the supply voltage by keeping the frequency constant. Due to this power increases more rapidly because power is directly proportional to the square of supply voltage [19]. DPM allows MPSoCs to minimize power and energy consumption by optimizing the dynamic power. We can reduce the frequency which saves a considerable amount of power but causes performance degradation in the multiprocessor. In the same way, as the supply voltage decreases the dynamic power is reduced almost four times but it has an overhead by reducing the supply voltage. An increase in circuit delay occurs so the circuit cannot operate at the same frequency. If we decrease the supply voltage and frequency the dynamic power decreases cubically but an increase in the execution period occurs linearly that degrading the performance of the chip [20].

DPM is a design technology that reconfigures the whole computing system dynamically in such a way that requested services can be provided with the minimum number of active CPUs with suitable performance levels [21]. Proposed a dynamic voltage and frequency scaling technique that is used to improve the performance and consider the load balancing issue in multi-cores of a processor. DVFS technique dynamically set the workload on the cores for this an irregular parallel divide and conquer algorithm is used to equally share the workload that reducing 31% of energy consumption at 400 MHz [22].

When task execution is interrupted in modern CPUs during the transition to a low-power state each low-power state σx is defined by its power consumption Pσx while the time and resource requirements of entering or leaving that state are represented as δx→x,δx→s,Es→x,andEs→s taking the sum of both the transition overheads of initial and final is necessary to get the total energy overhead that is linked with low-power state σx namely δx and Ex for comparison. Power usage is lower but the time and energy overhead are higher in the transition state to move the MPSoC into the low-power state as shown in Fig. 2. Where P is the processor’s power usage in its normal condition when no jobs are running while Pu is a specific idle state with a low transition overhead occurs. If the CPU is kept active throughout the idle interval the power consumption will be considered at the lowest speed. Various parameters define different low-power states. The two alternative state transitions represent a low-power state σ1 with moderate power usage and a rapid breakeven time while on other hand the low-power state δ2 reduces the power usage. The transition of state in low power mode occurs for short period due to the involvement of time and energy overheads [23]. Processors can achieve an active state rapidly when the transition’s energy overhead is minimal in such conditions the power state is known as the power-saving state.

TBe=max(δxEx−δxs∗PσxPu−Pδx)(4)

Figure 2: Two low-power states of an MPSoC

Eq. (4) represents the parameter Be known as the break-even time. Be refers to the minimum length of idle interval that must be provided to schedule and utilizes the sleep state δx efficiently. Be is the sum of the total time required to complete the state transition and the duration of idle time that is required to find a way to reduce the shifting energy [24]. The below Eq. (5) determines the BETsleep break-even time.

BETsleep=max(tpaEPa−Psleep∗tpaPidle−Psleep)(5)

The short circuit power is utilized when both PMOS and NMOS are on for a short period. Eq. (6) represents the transition of the state E0 and its power dissipation in running and the idle state as Pw and Ps [25].

Pw=Tbe=E0+Ps∗(Tbe−T0)(6)

In Eq. (7) (Tbe) breakeven time measures the length of the idle state of the CPU to optimize power.

Tbe=(E0−Ps∗T0Pa−Ps)(7)

DPM analyzes the run time behavior of the system to reduce the power consumption of the MPSoC. During the running state of an application a selective shut-down of the system components occur that are in the idle state increases the performance. When the CPU starts to transition the energy required for the state transition from sleep to idle and from idle to running is represented as Ei and its power dissipation at state 1 is denoted as Pa and at state 2 its Ps. For highperformance delay due to the state transition must be less than Tbe as shown in Eq. (8):

Tbe=max[(Ei−Ps∗T0Pa−Ps)∗T0](8)

Fig. 3 illustrates the behavior of MPSoC. On the left side, the MPSoC is running while on the other hand the device is in an idle state. The energy consumption on both ends is equal because of the break-even time in DPM-based techniques [26].

Figure 3: Task mapping in the system on chip working/idle state

Proposed an accuracy measurement model based on scheduling for medical imaging using a high-quality multiprocessor CPU and general processing unit (GPU) based computing environment to speed up the simulation rate and enhance the real-time performance [27]. Fig. 4 illustrates the optimization of power using the DPM technique that shows if the workload is not uniform on a system therefore the idle component of the system is considered [28].

Figure 4: Optimization of power with DPM

2.2 Predictive Techniques

A technique that measures the length of the next idle period of the processor using a simulation tool for real-time multiprocessor (STORM). The decision can be quickly made when the processor is in sleep mode and this policy doesn’t predict the length greater or less than the break-even period (Tbe). Introduced an exponential average scheme that predicts the next idle period length by considering an exponential average of the actual lengths of the previous idle period and the predicted one as shown in Eq. (9):

ln+1=b⋅ln+(1−b)⋅ln(9)

The new predicted values are represented as ln+1 and the last predicted value is denoted as ln. Where ln is the latest idle period while b’ is considered as an attenuation constant ranging between 0 to 1 [29]. The hardware architecture of Marvel X-Scale Intel PXA-270 MPSoC is widely used for predictive offline DPM techniques for both homogeneous and heterogeneous processors using STORM [30].

3  Problem Definition

The most critical concerns in multi-core embedded systems are the performance and life span of the chip. The task scheduling and switching of jobs from one core to another are one of the major issues in today’s MPSoC. Increasing power dissipation and energy utilization increases on chip temperature and resistance which reduces the life span of the chip. It also affects the reliability (hotspots, thermal cycles) as well as lowers the maximum speed for all battery powered devices, particularly in embedded systems that cause multiple performance and reliability issues. The key design difficulty in a task migrationbased system is an accurate forecast of energy, power, coolest core, utilization factor and the workload that needs to be relocated on an individual CPU. In multiprocessor systems task switching across various cores is a prevalent problem. Because the destination core may be in sleep mode there may be a delay while moving tasks from one core to another. Task migration is a technique for reducing power and energy while improving performance.

Power and energy optimization on multi-core systems are developed to address MPSoC dissipation concerns. The influence of tasks may be recognized when looking for the best solution task factors can be examined as tasks have an impact on each other. An energy-efficient task migration policy based on an EDF algorithm that optimizes energy while considering different configurations for migration of load is proposed in this research work. Normally, task schedules on each core are independent.

Problem 1: Energy-Efficient Multiprocessor Scheduling Technique (EEMS).

Consider a system having a periodic task set τ=(τ1,τ2,τ3,τ4,τ5….,τn) over midentical processors where many tasks in τ have a common deadline D and are ready at time 0. Each task τi ∈ τ is linked with CPU requirement that is approximately equal to power consumption function Pi(s) and CPU-cycles ci at the given CPU speed. The objective is to reduce the total energy consumption, of the CPU by scheduling for τi ∈ τ over midentical to complete before the deadline by allowing the suitable task migration among processors and core switching technique.

3.1 Proposed Energy-Efficient Multiprocessor Scheduling Algorithm with Task Migration

In this section, we have proposed EA-EDF an optimal algorithm for the EEMS problem. Since the power consumption Pi(s)/s is increasing at each cycle for every task τi ∈ τ in a given periodic task set, the CPU executes each task τi ∈ task set τ at some speed s. For this we have proposed EA-EDF an energy-efficient earliest deadline first scheduling algorithm based on the DPM technique for task migration policy and a guarantee on the energy consumption for MPSoC by monitoring the utilization factor. An energy-efficient policy for various configurations of cores to predict the power and energy profiles using both hardware & software architecture of the Intel PXA-270 MPSoC. Scheduling technique based on task migration reduces energy and produces energy-efficient results according to the utilization factor (ui). The proposed model uses an EA-EDF scheduler for energy and power optimization and is independently demonstrated to be the best-performing technique at run time.

The suggested policy operates with a set of parameters and generates effective outcomes for the given workload while also optimizing the chip’s lifespan and improving QoS by lowering energy consumption. The task set generator provides randomized task sets under a ready task queue that are being used to generate workloads under various constraints as well as the number of CPU cores specified by the user in extensible markup language (XML) based on the application given to STORM as an input. EA-EDF the scheduler can also have a set of prepared tasks that are scheduled according to the scheduling policy by considering suitable migration of task policy that produces energy and power profiles on the individual cores according to the given set of parameters in XML. A periodic task set τ=(τ1,τ2,τ3,τ4,τ5….,τn) and a utilization factor that is non-increasing (ui=ui+1for\;all\;i=1,≤i≤n,\;where\;ui=αipi), consider ui=∑j=0i=1for∀i=1,≤i≤n\; Let ϕ′ is a set of CPU with τ&m≤n & ca1,cb1,cc1,…cm, capacities ci≥ci+1for∀i=1,≤i≤m,τ can be scheduled on the uniform MPSoC that meets all the deadlines un≤cm;\;where\;ua≤ca\;for\;all\;k=1,2,.m; Overall utilization factor: ui=αipifor∀c=τi(ci,pi,,di) and n periodic task τ=(τ1,τ2,τ3,τ4,τ5….,τn) scheduled on a CPU the total ui is defined as uτ=∑k=1ncapa ≤ 1; for∀load a set of τ=τi(ca,pa,,da);\;for∀ n periodic task {τ1,τ2,τ3….τn} with random deadlines scheduled on a CPU, Load (τ) = max {Uτ }, Eq. (10) is used to measure the densities of MPSoCs as:

∂a=camin{da,Ta,}(10)

For the elaboration of the power and energy model Eq. (11) elaborates the sub-threshold leakage (Ist) that contains reverse bias junction current (Irbj).

Pleakage=vddIst+|Vbs|Irbj(11)

Dynamic energy per cycle is represented as Edyn as shown below in Eq. (12):

Edyn=Ceff⋅V2⋅f(12)

The total consumption of power is due to static and dynamic power that is used to represent the energy consumption of MPSoC cores. Static power is consumed while the CPU is on, whereas dynamic power is spent during calculation periods as shown in Eq. (13).

∫0tmax⁡p(t).dt,Pt=Pst+Pdy(13)

EPer−Cycle=f−1∗Lg∗vddIsubn+|Vbs|Ij(14)

The leakage energy per cycle is elaborated in Eq. (14). The delay per cycle is denoted as f−1 and the required energy for running state per cycle is Erunning=f−1∗Pon enhances when the frequency is low. Eq. (14) represents the total energy consumption per cycle that is represented as:

EtotalPer−Cycle=Edynamic+EPer−Cycle+Eon(15)

The proposed scheduling algorithm works for all real-time jobs in a running state using the system model represented in Fig. 5 allowing all new tasks in the ready state to go to the scheduling phase and execute on time. When compared to the prior strategies utilized in the earliest deadline first algorithm the scheduling methodology improves the chip’s overall working performance. When the tasks are ready to run the proposed scheduler verifies that the tasks are compatible with the scheduler and map them to cores based on the utilization factor (ui). In the proposed system model dynamic fixed priorities are assigned to the entire n periodic task τ={τa1,τb1,τc1…….πn} with deadline (da=pa).

Figure 5: System model

3.2 Core Configurations for EEMST

Scheduling and switching of τ jobs on various cores configurations are based on the utilization factor (ui). The amount of tasks in the running state requires a suitable task migration policy that defines the criteria for scheduling of readyτ on the core. When tasks are running the ui remains constant but as the number of tasks increases due to the concurrent processing of tasks on the multi-core processor’ui rapidly increases. For reducing the energy and power consumption optimal energy-efficient core configurations are defined for proper core switching and load balancing using task migration based on this policy. Various core configurations according to the ui utilization factors are mentioned below:

For∀τ→uτ=∑k=1ncapa≥(1%→9%),ui≥(9.1%→18%),

ui= 7%, P = 1 Core configuration {(1,), (2,),(3), (4),(5),(6),(7),(8)}

ui≥(18.1%→27%),ui= 14%, P = 2 Core configuration {(1, 3),(4, 7),(2, 5), (6, 8)}

ui≥(27.1→36%),ui= 24%, P = 3 Core configuration {(1, 3, 5), (4, 7, 8), (1, 2, 6)}

ui≥(36.1%→45%),ui= 31%, P = 4 Core configuration {(1, 3, 57),(2, 4, 6, 8)}

ui≥(45.1→54%),ui= 40%, P = 5 Core configuration {(1, 2, 3, 7, 8),(1, 2, 4, 5, 6)}

ui≥(54.1→62.5%),ui= 47%, P = 6 Core configuration {(1, 2, 3, 7, 8), (1, 2, 4, 5, 6)}

ui≥(54.1→62.5%),ui= 67%, P = 7 Core configuration {(1, 2, 3, 5, 7, 8), (1, 2, 4, 5, 6, 8)}

For∀τ→uτ=∑k=1ncapa>(62.6%), P  = 8 Core configuration {(1, 2, 3, 4, 5, 6, 7 and 8)}

4  Simulation and Results

A simulation tool for real-time multiprocessor (STORM) is used to perform the experimental valuation on a homogeneous multiprocessor system and illustrate the importance due to advancements in semiconductor technology that increases the power density of a multiprocessor with the increase in demand of real-time applications based on a complex circuitry. STORM receives ready tasks from the XML input file as shown in the system model and schedules it on the hardware architecture of Marvel X-Scale Intel PXA-270 MPSoC by considering the suitable core configuration with DPM capabilities. Core configurations are chosen based on the utilization factor whereas the power model for all cores is the same.

Intel PXA-270 processor’s power model is used for each core in the test scenarios because the Intel-PXA270 offers seven power states. We tested the behavior of our proposed method by running all tasks τ till the arrival of the worst-case execution time (WCET). For this we created a task set and stored the data in an XML input file. The task set’s size ranges from 5 to14 tasks with each task’s period falling within [0.15 ms, 14 ms] as indicated. We executed the task set mentioned in Tab. 2 on the Intel-PXA270 MPSoC processing platforms. The XML input contains, starts time, WCET, period, priority, deadline, hardware architecture, number of cores and proposed EAEDF scheduling mechanism. The proposed scheduling algorithm using task migration ability is implemented on homogeneous multi-core architecture that has more than one core and shares the same architecture and microarchitecture integrated on a single chip.

Tab. 3 represents various power consumption states of Intel PXA-270 over multiple operating frequencies respectively by considering various states of running, idle and sleep. It also represents the current of the Intel PXA-270 CPU on various frequencies. All hardware characteristics from the Intel-PXA270 processor are employed in our tests to achieve better performance and minimize delays in the execution process. The proposed EA-EDF migrates tasks to a core that is physically away from the core with the highest power consumption and temperature. This section presents the evaluation of our improved energy-efficient scheduling algorithm for the optimization of energy and power. The proposed energy aware-EDF-based scheduler gives improved results and more energy optimization as compared to previous techniques. The Intel PXA-270 MPSoC is used to measure CPU energy usage. According to the proposed algorithm the job with the earliest scheduling deadline is prioritized when the periodic task set is examined the strategy works well that’s why the context switch is valuable.

Considering the same experimental characteristics such as deadlines and consumption. The simulation indicates that the STORM simulator performance criteria are equivalent to the real platform values. Tab. 4 represents the comparison of experimental results of energy consumption on 520 MHz at various ui = 6%, 10%, 20%, 36%, 55%, 60%, 62.5% and ui ≥ 63% when all the cores are running.

Tabs. 5 and 6 represents the comparison of proposed EA-EDF at various workloads ui as compared to other currently used techniques on 520 and 624 MHz. The simulation results show that the proposed EA-EDF is more energy-efficient as compared to other techniques at lower utilization and gives 4.7% & 4.3% more energy efficient results on 10%, 27% & 36% ui as compared to GEDF, RT-DPM, TBP, PDTM on 520 & 624 MHz in terms of power & energy. The Intel PXA-270 MPSoC is used to measure CPU energy usage. The proposed EA-EDF algorithm prioritized the job with the earliest scheduling deadline.

The comparison of energy consumption at various utilization factors ui on 520 MHz as shown in Fig. 6 illustrates the results of the proposed energy-aware EA-EDF scheduler that behaves similar to GEDF and PDTM for∀τ→uτ=>62.5%, but optimizes 4.7% of the overall energy at for∀τ→uτ<62.5%.

Figure 6: Comparison of energy consumption ui at 520 MHz

Fig. 7 represents the power consumption on various cores of MPSoCs with the increasing number of tasks using EA-EDF atui=36. Only 4 cores Core1,Core2,Core3&Core4, are running by intelligently migrating and switching the load towards the low energy cores before ui meeting the threshold. The features of the task set include the software architecture, start time and worst-case execution time, period, priority and deadline. While hardware architecture contains the no of cores and the proposed scheduling algorithm that is defined in the XML file for achieving the experimental results in Fig. 8.

Figure 7: CPU power consumption at various CPU cores under ui = 36% utilization at 520 MHz

Figure 8: Comparison of energy consumption ui at 624 MHz

The comparison of energy utilization of EA-EDF at ui = 6%, 10%, 20%, 36%, 55%, 60%, 62.5% and ui ≥ 63% when all the cores are in running state on 624 MHz as shown in Fig. 8. The proposed energy-aware EA-EDF scheduler had an impact on the system performance and enhances the life span of the chip by optimizing the average energy of 4.7% for∀τ→uτ<36.5% by giving smoother energy utilization patterns between individual cores for a set of periodic R-T tasks with a common deadline.

5  Conclusions

This research paper targets energy-efficient multiprocessor scheduling problems over τi ∈ τ over midentical homogeneous processors for a set of periodic R-T tasks with a common deadline for energy optimization for this problem we have developed an accurate energy optimization-based task migration policy that will be used to calculate the load of the destination core. For the selection of cores, various configurations have been proposed in terms of energy efficiency and utilization factor that enables a processor to reduce power and energy consumption by adopting a suitable task migration policy. An increase in power utilization reduces the life span of the chip and has become an integral chip design issue for the battery-operated multiprocessor system. The proposed EA-EDF model is based on an optimal scheduling policy for accurate migration of tasks without missing their deadlines by considering the hardware architecture of Intel PXA-270 MPSoC in the STORM simulator. The objective is to increase the performance of the homogenous multi-core systems by gradually decreasing the thermal cycles, power and energy consumption. whereas Less number of running cores are combined in configuration. Proposed EA-EDF algorithm enables load balancing and allows processes to run on various cores at various times as decided by the scheduler using periodic task sets with implicit deadlines on homogeneous multiprocessor platforms. In future work, our algorithm can be used for heterogeneous multiprocessor platforms apart from this the proposed technique gives us more efficient results by optimizing 4.3%–4.7% energy on a utilization of 6%, 36% & 46% at 520 & 624 MHz operating frequency when compared to previously deployed GEDF, TLP, PDTM, RT-DPM TBP energy-based optimization techniques.

Acknowledgement: The researchers wish to extend their sincere gratitude to the Deanship of Scientific Research at the Islamic University of Madinah for the support provided to the Post-Publishing Program1.

Funding Statement: The authors received no specific funding for this study.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.