Sage Journals: Discover world-class research

Abstract

Due to the importance of power dissipation in the wireless sensor networks and embedded systems, real-time scheduling has been studied in terms of various optimization problems. Real-time tasks that synchronize to enforce mutually exclusive access to the shared resources could be blocked by lower priority tasks. While dynamic voltage scaling (DVS) is known to reduce dynamic power consumption, it causes increased blocking time due to lower priority tasks that prolong the interval over which a computation is carried out. Additionally, processor slowdown to increase execution time implies greater leakage energy consumption. In this paper, a leakage-controlled method is proposed, which decreases both priority inversion and power consumption. Based on priority ceiling protocol (PCP) and a graph reduction technique, this method can decrease more energy consumption and avoid priority inversion for real-time tasks.

1. Introduction

Power consumption of IC devices can be briefly classified into (1) static power consumption which consists of the power used when the transistor is not in the switching process and (2) dynamic power consumption which consists of the power used to charge the load capacitance and logic states of the device. In the last two decades, complementary metal-oxide semiconductor (CMOS) has emerged as a dominant technology due to its low static power dissipation. Leakage power in CMOS circuits is due to subthreshold current that flows through the transistors and contributes to a significant portion of the total power consumption. Figure 1 illustrates the trend at lower nodes that leakage power becomes almost comparable to active power. For 90 nm and below technologies, leakage is the main factor which dominates over the dynamic power and contributes to almost or more than 50% of total power dissipation.

Figure 1

Leakage versus dynamic power trend [1, 19, 20].

There have been multiple techniques used in the past to reduce the power dissipation of embedded systems and have been implemented successfully through the different levels of design abstraction. At the system level, there are two primary ways to reduce power consumption: processor slowdown and shutdown. Slowdown techniques such as dynamic voltage scaling (DVS) are preferred due to the assumptions of low leakage power dissipation and quadratic dependence of power on voltage level. However, energy savings based on DVS come at the cost of increased execution time, which implies greater leakage power consumption. With the repaid increase in device leakage power due to the scaling down of technology nodes, the leakage power has been a major concern in the system design. Device scheduling is the key factor to have the best tradeoff between the power and characteristics such as resource sharing and system utilization.

Many previous works have addressed DVS based on performance requirements to minimize the dynamic power consumption [1–13]. While techniques to optimize the total static and dynamic power consumption have been proposed, they are still based on the ideal system environment without considering resource synchronization problem. Recently, researchers have started exploring energy-efficient scheduling with the considerations of leakage current [4–6, 14–16]. To reduce the leakage power, a system might be turned off (to enter a dormant mode) when not used. For periodic real-time tasks, Jejurikar et al. [6] and Lee et al. [14] proposed energy-efficient scheduling on a uniprocessor by procrastination scheduling to decide when to turn off the system. Jejurikar and Gupta [5, 15] further considered real-time tasks that might complete earlier than its worst-case execution time (WCET) by extending the algorithms presented in [6]. For DVS processor with aperiodic real-time tasks, Irani et al. [4] proposed a 3-approximation algorithm to minimize leakage current on the uniprocessor with continuous available speeds. It guarantees to derive a solution with the energy consumption at most three times energy consumption of an optimal solution. For aperiodic real-time tasks without DVS, Baptiste [17] developed an algorithm that minimized the energy consumption resulting from the leakage current and the overhead to turn on/off the system. In the periodic real-time task systems, Quan et al. [13] and Niu and Quan [16] proposed a leakage-aware method that expands all the jobs in the hyperperiod of the given tasks and calculates their latest starting time on the fly. The computation overheads produced by their algorithms might be too much to be applied in an online scheduling. In addition to considering the leakage current of the processor, leakage power consumption produced by system devices was also considered in [18] as well as their preemption control [9]. The objectives of the above methods are proposed to reduce the system power and do not consider other advantages or purposes of leakage-aware techniques. For example, one can shut down a processor before the beginning of a lower priority task that intents to lock a shared resource required earlier by a higher priority task.

Work-demanded analysis has been extensively investigated in real-time systems in which sporadic tasks are jointly scheduled with periodic tasks [3, 7–9, 11, 12, 15, 21]. The purpose of work-demanded analysis is to improve the response times of aperiodic tasks or to increase their acceptance ratio. It also benefits the power-saving on both DVS and leakage-aware methods. Pillai and Shin [12] proposed a cycle-conserving rate-monotonic (ccRM) scheduling scheme that contains offline and online algorithms. The offline algorithm computes the worst-case response time of each task and derives the maximum speed needed to meet all task deadlines. It recomputes the task utilization by comparing the actual time for completed tasks with WCET schedule. In other words, when a task completes early, they have to compare the used actual processor cycles to a precomputed worst-case execution time schedule. This WCET schedule is also called canonical schedule [22] whose length could be the least common multiplier of task periods. ccRM is a conservative method, as it only considers possible slack time before the next task arrival (NTA) of current job. Gruian proposed a DVS method for offline task stretching and online slack distribution [3]. The offline part of this method consists of two separate techniques. One focuses on the intratask stochastic voltage scheduling that employs a task-execution length probability function. The second technique computes stretching factors by using a response time analysis. It is similar to Pillar and Shin's offline technique, but instead of adopting a stretching factor for all tasks before NTA, Gruian assigns different stretching factor to the individual task within the longest task period. Kim et al. [8] proposed a greedy online algorithm called the low-power work-demand analysis (lpWDA) that derives slack from low-priority tasks, as opposed to the method in [3, 12] that gains slack time from high-priority tasks. This algorithm also balances the gap in voltage levels between high-priority and low-priority tasks. Its analysis interval limited by the longest of task periods is longer than NTA. Thus, lpWDA gains more energy saving than the previous rate-monotonic (RM) DVS schemes applying NTA. Therefore, these slack reclamation methods are applicable not only to increase energy-saving and schedulability but also to the new methods on different scheduling criteria. For example, slack time can also be applied to avoid priority inversion incurred by task synchronization.

Though power-aware real-time scheduling problems are well studied, relatively few works of them considered task synchronization problems. Most embedded real-time applications have to share the resources such as I/O devices or database and mutually exclusive access to these shared objects. Concurrent access to shared data may result in data inconsistency. Mechanisms that force tasks to execute in a certain order so that data consistency is maintained are known as synchronization protocols. If a lower priority job has access to a resource, a higher priority job requesting the resource is blocked and can miss its deadline. Therefore, real-time task synchronization is developed to achieve data consistency with as little loss in schedulability as possible. Priority ceiling protocol (PCP) [23, 24] is a synchronization protocol for shared resources to avoid deadlock and unbounded priority inversion due to wrong nesting of critical sections. In PCP, each resource is assigned a priority ceiling, which is a priority equal to the highest priority of any task which may lock the resource. PCP has the interesting and very useful property that no task can be blocked for longer than the duration of the longest critical section of any lower-priority task. The recent related work is an extension of PCP in frequency inheritance. Zang and Chanson [25] proposed a dual-speed (DS) algorithm: one is for the execution of a task when it is not blocked, and the other is adopted to execute the task in the critical section when it is blocked. Jejurikar and Gupta [26] compute two slowdown factors, which can be classified into static slowdown, computed offline based on task properties, and dynamic slowdown, computed using online task execution information. Chen and Kuo [2] proposed a DVS method using frequency locking concept which can be used to render energy-efficient to the existing real-time task synchronization protocols. However, DVS may decelerate the blocking tasks in the critical section and receive additional priority inversion which increases the difficulties of predictability of the schedule.

In this paper, we propose a blocking-aware task synchronization method, which decreases both priority inversion and energy-consumption of the processor. Graph theories have been applied in the scheduling which transforms an existing resource allocation graph (bipartite graph) to a weighted-directed graph (WDG). Our idea is based on PCP and available slack time in the schedule. We use WDG to analyze and present the available slack time, preemption points, and the lengths of priority inversions. The timing and its duration for switching processor status can be computed by traversing the WDG in which lower-priority tasks have intention to lock on some resources. The proposed method can utilize the existing work-demanded analysis methods such as lpWDA [8] and Gruian's methods [3] to squeeze additional slack time from lower priority tasks so as to further decrease priority inversion.

The remainder of this paper is organized as follows. Section 2 explains our system model. The motivations and basic idea on the task synchronization method are proposed in Section 3. Section 4 describes the proposed algorithm and its example. We present the performance evaluation in Section 5. Section 6 gives conclusions and their future work.

2. System Model

2.1. Task Model

This paper studies periodic real-time tasks that are independent during the runtime. Let 𝒯 be the set of input periodic tasks and n denotes the number of tasks. Each task $τ_{i}$ is an infinite sequence of task instances, referred to as jobs. The jth invocation (job) of task $τ_{i}$ is denoted as $J_{i, j}$ whose actual start time is denoted as $S (J_{i, j})$ . A three-tuple $τ_{i}$ = ${T_{i}, D_{i}, C_{i}}$ represents each task, where $T_{i}$ is the period of the task, $D_{i}$ is the relative deadline with $D_{i} = T_{i}$ , and $C_{i}$ denotes the worst-case execution time (WCET). The period length of task $T_{i}$ is unique in order to have each task a unique priority index in the rate-monotonic (RM) scheduling. Notation $s_{i, j}$ denotes the available slack time for job $J_{i, j}$ .

2.2. Power Model

All tasks are scheduled on a single processor which supports two modes: dormant mode and active mode. When the processor is switched to the dormant mode, the power-consumption of the processor is assumed $S_{dorm} = 0$ by scaling the static power consumption [27], while the system clock and chipset retain necessary functions to support motoring and waking up processor properly. When executing the jobs, the processor has to be in the active mode with speed $S_{active}$ . A processor shutdown and wakeup have a higher overhead than the inherent energy/delay cost of turning on the processor. The processor loses the cache and register contents, when it switched to the deepest dormant mode. Therefore, prior to shutdown, all statuses of processor must be saved or flushed to main memory, resulting in an additional overhead. On wakeup, components such as data and instruction caches, data and translation look aside buffers, (TLBs) have to be initialized. This results in extra memory accesses and hence additional energy consumption. Let $E_{sw}$ and $t_{sw}$ denote energy and time overhead, respectively, for switching from dormant mode to active mode. Based on the idle power consumption, when processor is idle in the active mode, the processor executes NOP instruction at processor speed $S_{idle}$ for low-power consumption referred to as $P (S_{idle})$ . When processor is idle and the idle interval is longer than break-even time $E_{sw} / P (S_{idle})$ , turning it to the dormant mode is worthwhile.

2.3. Resource Access Model

We assume that semaphores are used for task synchronization. All tasks are assumed to be preemptive, while the access to the nonpreemptive shared resources must be serialized. Therefore, a task can be blocked by lower priority tasks. When a task has been granted access to a shared resource, it is said to be executing in its critical section [24]. The kth critical section of task $τ_{i}$ is denoted as $z_{i, k}$ which is properly nested. Each task knows the relative start time to access the required shared resources and their required WCETs. With the given information in [28], we can compute the maximum blocking time for a task. In the different resource synchronization protocol such as PCP [23], each job might suffer from a different amount of blocking time from lower-priority task due to access conflict. The goal of this paper is to propose a leakage-aware task synchronization protocol based on PCP, which reduces the priority inversion and energy-consumption of the processor. We propose a data structure called weighted directed graph (WDG) which expresses possible priority inversion online. By traversing WDG, we can postpone the intention to locking on the resources invoked by lower-priority tasks and therefore minimize the priority inversions.

3. Motivating Example

Suppose that we have three jobs $J_{1}$ , $J_{2}$ , and $J_{3}$ and three shared data structures protected by the binary semaphores $z_{0}$ , $z_{1}$ , and $z_{2}$ in the system. In accordance with PCP, the sequence of events is depicted in Figure 2(a). A line at a lower level indicates that the corresponding job is blocked or preempted by a higher-priority job while the processor mode is active. A line raised to a higher level denotes that the job is executing, and the absence of a line denotes that the job has not yet been initiated or has completed. A bold line at low level denotes that the processor has been switched in dormant mode. Suppose the following. (i)

At time $t_{0}$ , $J_{3}$ is initiated and locks semaphore $z_{2}$ at $t_{0}^{'}$ .

(ii)

At time $t_{1}$ , $J_{2}$ is initiated and preempts $J_{3}$ .

(iii)

At time $t_{2}$ , $J_{2}$ cannot lock $z_{2}$ and $J_{3}$ inherits the priority of job $J_{2}$ and resumes execution.

(iv)

At time $t_{3}$ , $J_{3}$ successfully enters its nested critical section $z_{1}$ .

(v)

At time $t_{4}$ , $J_{1}$ preempts $J_{3}$ within $z_{1}$ and executes its noncritical section code.

(vi)

At time $t_{5}$ , $J_{1}$ attempts to enter its critical section $z_{0}$ and is blocked by $J_{3}$ due to priority ceiling.

(vii)

At time $t_{6}$ , $J_{3}$ exits its critical section $z_{1}$ and $J_{1}$ is awakened.

(viii)

At time $t_{7}$ , $J_{1}$ completes its execution and $J_{3}$ resumes its execution of $z_{2}$ due to its inherited priority.

(ix)

At time $t_{8}$ , $J_{3}$ exits its critical section $z_{2}$ and returns to its original priority, and $J_{2}$ is awakened.

Figure 2

The task synchronization of (a) primitive PCP and (b) latency locking method.

The blocking intervals introduced by primitive PCP are [ $t_{2}$ , $t_{4}$ ], [ $t_{5}$ , $t_{6}$ ], and [ $t_{7}$ , $t_{8}$ ].

This study is motivated by the significant priority inversion that is incurred from task synchronization. When the available static slack time (unused time in the WCET schedule) or dynamic slack (occurred in the early-completed task) is larger than break-even time, the lower-priority task intent to lock a semaphore can be postponed until the start time of a higher-priority task. A practical approach is to postpone the task execution by switching processor to dormant mode. During the sleeping time, system still has awareness of the arrival of other jobs and awakes processor at proper time. The example in Figure 2(b) postpones the request of lower-priority task intent to lock a semaphore. At time $t_{1}$ , $J_{3}$ has available slack in interval [ $t_{11}$ , $t_{12}$ ] with length longer than break-even time. When a system is conscious that $J_{3}$ has intent to lock $z_{2}$ , it computes the upcoming start time of higher priority tasks that might be blocked by $J_{3}$ according to PCP. In the example, only $J_{2}$ could be blocked by $J_{3}$ due to $z_{1}$ and $z_{2}$ , and the length of interval [ $t_{1}$ , $t_{2}$ ] is less than the available slack, that is, [ $t_{10}$ , $t_{11}$ ] in Figure 2(a). Therefore, processor can be switched to dormant mode at time $t_{1}$ until the start time of $J_{2}$ . At time $t_{2}$ , processor becomes active and $J_{2}$ preempts $J_{3}$ such that $J_{3}$ is still unable to lock $z_{2}$ , and thus $J_{2}$ could lock $z_{2}$ at time $t_{3}$ . Additionally, the $J_{2}$ 's noncritical section also is not blocked by $J_{3}$ because $J_{3}$ cannot inherit the priority of $J_{2}$ due to failure in locking $z_{2}$ at $t_{1}$ . Comparing to the result of Figure 2(a), this idea can reduce (maximum) priority inversion of intervals [ $t_{2}$ , $t_{4}$ ], [ $t_{5}$ , $t_{6}$ ], and [ $t_{7}$ , $t_{8}$ ]. Table 1 presents the length of blocking time in a systematic way [28]. It lists only the nonzero elements; all the other elements are zero, since jobs are not blocked by higher priority jobs. The elements labeled with “*” in this table are zero as well.

Table 1

Example illustrating the computation of blocking times.

	Directly blocked by		Priority-inher blocked by
	$J_{2}$	$J_{3}$	$J_{2}$	$J_{3}$	max
$J_{1}$		3			3
$J_{2}$	*	4	*	3	4
$J_{3}$	*	*	*	*	*

4. Latency Locking PCP

In this study, PCP is extended with the concept of latency locking, referred to as LL-PCP. The idea is to perform preanalysis of possible priority inversion and available slack time in the schedule. LL-PCP derives the best timing and duration for switching processor to dormant mode and thus reduces priority inversion. To understand and control the sequence of intent to lock resources, tasks are organized as a weighted directed graph (WDG) reduced from the resource allocation bipartite graph in [28]. In the bipartite graph, each indirect edge is labeled with the time required to access the resources. Let $G = (U, V, E)$ denote a bipartite graph whose partition of vertices has two subsets U and $V . E$ denotes the set of edges of G, and U denotes a task set 𝒯. Let WDG( $𝒯, A$ ) denote a weighted directed graph whose vertices in $𝒯 \subseteq U$ are arranged according to their task indices. For each edge $e_{u, v}$ $\in E$ , $τ_{u} \in U$ , and $z_{v} \in V$ , the set of arcs A in WDG are generated as follows.

Step 1.

For any pair of vertices $τ_{x}$ , $τ_{y} \in U$ , and $x > y$ , a solid arc $a (x, y) \in A$ is directed from $τ_{x}$ to $τ_{y}$ if there exist two or more edges $e_{x, v}$ and $e_{y, v}$ in G, where $z_{v} \in V$ .

Step 2.

For any pair of vertices $τ_{x}$ , $τ_{w} \in U$ , and $x > w$ , a dotted arc $a (x, w) \in A$ is directed from $τ_{x}$ to $τ_{w}$ if there exists a vertex $τ_{y} \in U$ , $w > y$ , and $τ_{x}$ and $τ_{y}$ satisfy Step 1.

Step 3.

In WDG, for any pair of vertices with multiple arcs, eliminate the dotted arcs having the same blocking time as that of one of their solid arcs.

Different from the bipartite graph, each vertex in the WDG corresponds to a task but resource. A task $τ_{L}$ directly blocking a higher-priority task $τ_{H}$ is represented by a solid arc $a (τ_{L}, τ_{H})$ from task vertex $τ_{L}$ to $τ_{H}$ , while an indirect block is represented by a dotted arc. In Figure 3(a), the bipartite graph is derived from Figure 2(a) and can be reduced to the WDG in Figure 3(b). The maximum priority inversion of $J_{2}$ is composed of a direct and indirect blocking incurred by $J_{3}$ . The values of these two arcs can be combined because the indirect blocking is due to priority inheritance of job $J_{1}$ . We may label each arc by a 3-tuple; the first element of each 3-tuple gives the actual starting time of higher-priority tasks and is defined as $S (τ_{H})$ , while the second element gives the locking time of $τ_{L}$ on semaphore z and is denoted as $L (τ_{L}, z)$ . The last element specifies the duration of the maximum priority inversion and is denoted as $I (τ_{L}, τ_{H})$ . The first two elements are updated during runtime while the third element is derived according to PCP. The example of the 3-tuple labels is illustrated in Figure 3(b). In accordance with the parameters, the proposed evaluation functions are defined as shown in Algorithm 1.

Algorithm 1: LL-PCP algorithm.

Input: a set of task T, a set of resources R.

(Offline part)

(1) Reduce bipartite graph to WDG (T, A);

(2) Compute the value of I ( $τ_{L}$ , $τ_{H}$ ) with respect to each arc

a ( $τ_{L}$ , $τ_{H}$ ) in WDG.

(Online part)

On arrival of a job $J_{i}$

(3) Identify a set of tasks $T_{H}$ containing higher

priority task $τ_{H}$ than that of $J_{i}$ ;

(4) Update the value of REW ( $τ_{i}$ , $τ_{H}$ ) for each

arc a( $τ_{i}$ , $τ_{H}$ ) in WDG and $τ_{H} \in T_{H}$ ;

(5) Construct a set $A_{i}^{'}$ of outgoing arcs of $τ_{i}$ ,

$A_{i}^{'} = {a (τ_{i}, τ_{H}) ∣ α_{i, H} \geq \frac{E_{SW}}{P (S_{idle})}}$

(6) Compute the static available slack $s_{H}$ for

each job in $T_{H}$ ;

(7) Compare each $s_{H}$ to the corresponding α

value of the arcs in $A_{i}^{'}$ ;

(8) Construct an arc set $A_{i}^{''} \subset A_{i}^{'}$ where $A_{i}^{''}$

$= {a (τ_{i}$ , $τ_{x}) ∣ s_{H} \geq α_{x, i}$ , a( $τ_{i}$ , $τ_{x}$ ) $\in A_{i}^{'}$ and $τ_{x} \in T_{H}}$ ;

(9) Search for an arc a( $τ_{i}$ , $τ_{x}$ ) in $A_{i}^{''}$ with the

maximum value of REW( $τ_{i}$ , $τ_{x}$ ) where $τ_{x} \in T_{H}$ ;

(10) IF the processor is in active mode THEN

switch the processor to dormant mode until time

$S (τ_{x})$ ;

On the beginning of one of the intervals in ${E S I}_{L, x}$

(11) Switching the processor to $S_{dorm}$ until the

end of the interval;

On turning the processor to the active mode at time t

(12) Schedule the highest priority job in the

ready queue; On early-completion of a job at time t;

(13) Compute dynamic slack time due to early completion;

Figure 3

Graph reduction from (a) bipartite graph to (b) WDG.

Definition 1 (expected sleeping interval, ESI).

In order to prevent job $J_{L}$ from blocking job $J_{H}$ , the expected sleep interval for $J_{L}$ is defined as

\begin{matrix} ES I_{L, H} = [L (J_{L}, z), S (J_{H})) - ⋃_{\forall τ_{λ} \in 𝔍, λ < L} {N C}_{λ}^{[L, H]}, \end{matrix}

(1)

where

{N C}_{λ}^{[L, H]}

denotes the set of noncritical-section intervals of job

J_{λ}

in interval

[L (J_{L}, z), S (J_{H}))

. The length of

{ESI}_{L, H}

is denoted as

\begin{matrix} α_{L, H} = S (J_{H}) - L (J_{L}, z) - \sum_{\forall τ_{λ} \in I, λ < L} | {N C}_{λ}^{[L, H]} | . \end{matrix}

(2)

Definition 2 (reduction of priority inversion time, RPI).

The expected reduction of priority inversion time due to the processor sleeping in the ${ESI}_{L, H}$ is defined. The value of RPI is derived from

\begin{matrix} β_{L, H} = I (τ_{L}, τ_{H}) - α_{L, H} . \end{matrix}

(3)

According to (1), (2), and (3), we define a reward function for each arc in WDG.

Definition 3.

A reward function for each arc in WDG is defined as

\begin{matrix} R E W (τ_{L}, τ_{H}) = \frac{β_{L, H}}{α_{L, H}} . \end{matrix}

(4)

The reward for an arc referred to the reduction of priority inversion time if the processor is switched to sleep during the interval ${ESI}_{L, H}$ . Whenever a new job $J_{i}$ arrives, the value of $R E W (τ_{L}, τ_{i})$ with respect to each arc is refreshed. The larger the value of $R E W$ , the longer the priority inversion in the schedule will be avoided. For example, in Figure 3(b), the values of $α_{3, 1}$ and $α_{3, 2}$ are, respectively, $t_{4} - t_{0}^{'}$ and $t_{1} - t_{0}^{'}$ , and the values of $β_{3, 1}$ and $β_{3, 2}$ are, respectively, $3 - (t_{4} - t_{0}^{'})$ and $7 - (t_{1} - t_{0}^{'})$ . In accordance with (4), the values of $R E W (3,1)$ and $R E W (3,2)$ are $(3 - (t_{4} - t_{0}^{'})) / (t_{4} - t_{0}^{'})$ and $(7 - (t_{1} - t_{0}^{'})) / (t_{1} - t_{0}^{'})$ . Assuming that available slack for $τ_{3}$ is larger than the value of $(t_{1} - t_{0}^{'})$ , the processor decides to sleep in the duration of $[t_{0}^{'}, t_{1}]$ .

By updating the information of arcs in WDG during runtime, we can traverse the vertices of WDG by following the current job and make decisions on switching the processor to active or dormant mode. Additionally, the values of ESI can be updated on the fly according to the dynamic slack time due to early completion of a job. Also, the arcs corresponding to the unused resources due to early completion can be removed from this job vertex.

5. Experimental Results

To evaluate the effectiveness of the proposed scheduling method, we implemented the following techniques and derived their energy consumption. (i)

Primitive PCP: all tasks are executed at maximum speed and use PCP for accessing the shared resources [23].

(ii)

P-PCP: all tasks are scheduled using procrastination technique under fixed priority [5].

(iii)

DVS-PCP: all tasks are scheduled according to the RM-DVS scheduling based on the slowdown factor proposed in [26].

(iv)

LL-PCP: all tasks are scheduled according to the proposed method.

We use the processor power model in [5] and consider energy overhead of shutdown required by on-chip cache. With an energy cost of 13 nJ [15] per memory write, the cost of flushing the data cache is computed as 85 uJ. On wakeup, we assume a cost of 98uJ total energy overhead. Additionally, we assume that cache energy overhead to actual charging of circuit logic requires 300 uJ, and the total cost is $85 + 98 + 300 = 483 uJ$ . In the experiment, we consider 2000 task sets, and each set contains up to 12 randomly generated tasks. Each task was assigned a random period and WCET in the range $[10 ms, 125 ms]$ and $[0.5 ms, 10 ms]$ , respectively, to reflect real life real-time applications. A set of shared resources and its mapping to a set of tasks are generated randomly. For simplicity, the number of resources is less than or equal to the number of tasks of the corresponding task set and all resources are mutually exclusive accessed by those tasks.

In Figure 4, the comparison of the energy consumption of the techniques is as a function of the number of task, and the energy consumption of all techniques is normalized to LL-PCP. As the number of task increases, primitive PCP starts consuming more energy than other methods. When the number of tasks is greater than four, DVS-PCP starts consuming more energy due to the dominance of leakage and speed switching. The leakage is derived from frequent priority inversion due to increased resource contentions. We see that the LL-PCP results in an up to an additional 4.5% gains over DVS-PCP when the number of tasks is not less than 10.

Figure 4

Comparison of energy consumption in the different number of tasks, utilization = 50%.

In Figure 5, the energy consumption of the techniques is as a function of the processor utilization at maximum speed. When the processor utilization increases, DVS technique starts consuming more energy. As the processor utilization is low, DVS-PCP presents significant energy savings. There are idle intervals in these techniques and could be utilized by applying dynamic reclamation methods [3, 8, 12] for more energy gains. These idle intervals can be used for processor shutdown to compare the benefits of our procrastination scheme. Both P-PCP and LL-PCP are procrastination scheme and their energy consumptions outperform that of preemptive PCP through shutdown. Additionally, P-PCP and LL-PCP are not sensitive to the change in processor utilization because they are leakage-aware methods. We see that LL-PCP leads to up to 6% energy gains over P-PCP and an average of 5% energy savings compared to DVS-PCP.

Figure 5

Comparison of energy consumption in the different processor utilization, 6 tasks.

Figure 6 shows the comparison of the length of priority inversion normalized to LL-PCP, as a function of the number of tasks. In consideration of task synchronization, priority inversions directly affect the feasibility of real-time scheduling because they usually postpone actual finish time of higher-priority tasks and would result in deadline misses. When the number of tasks increases, primitive PCP, P-PCP, and DVS-PCP result in increasing priority inversion longer than that of LL-PCP. The main reason is that the number of increases of nonpreemptive resources and access conflictions among the tasks would produce additional short idle periods and switching time that also hamper timely completion of the tasks. LL-PCP reduces an additional 4.8% priority inversion over DVS-PCP at 12 tasks in each task set.

Figure 6

Comparison of priority inversion in the different number of tasks, utilization = 50%.

The comparison of the length of priority inversion of the techniques is shown in Figure 7, as a function of the processor utilization at maximum speed. As the processor utilization increases, preemptive PCP, P-PCP, and DVS-PCP result in less increasing priority inversion than that of LL-PCP. The reason is that, in accordance with (4), when the value of $α_{L, H}$ increases and $β_{L, H}$ remains unchanged, the returned values of their $R E W$ function also increases. Obviously, increasing the WCET of each task does not affect the length of priority inversion (i.e., the value of $β_{L, H}$ ) but increases the value of $α_{L, H}$ , and therefore the value of $R E W (\cdot, \cdot)$ is increased. In accordance with our algorithm, $R E W (\cdot, \cdot)$ affects LL-PCP not to switch lower-priority task to dormant mode and therefore produces some priority inversion. However, as the processor utilization is low, there are idle intervals in the schedules and they can be used by LL-PCP to reduce priority inversion. Nevertheless, LL-PCP still outperforms other techniques at any utilization.

Figure 7

Comparison of priority inversion in the different processor utilization, 6 tasks.

The average performance of the abovementioned methods is presented in Table 2 as the setting to the range of processor utilization and the number of tasks in each task set is [10%~90%] and 6, respectively. The proposed method still outperforms primitive PCP, P-PCP, and DVS-PCP with regard to both energy consumption and priority inversion.

Table 2

Comparisons of average energy consumption and priority inversion.

	Primitive PCP	P-PCP	DVS-PCP	LL-PCP
Energy consumption	100%	87.7%	86%	82%
Priority inversion	100%	97.2%	86.8%	84%

The proposed method may have additional offline processing time comparing to other methods. With respect to time complexity, the proposed method takes $O (n^{2})$ time during the offline processing because of the required data structures of WDG where n denotes the number of tasks. However, other methods that required analytical function to predict blocking time have to construct a blocking time table similar to Table 1 which also takes $O (n^{2})$ time. When using the existing slack-reclamation methods such as the algorithms in [3, 8, 22], the time complexities of the online parts of primitive PCP, P-PCP, DVS-PCP, and LL-PCP require $O (n \log n)$ time. In LL-PCP, the time complexities of each step are discussed as follows. Step (3) in the algorithm searching for the tasks with respect to job $J_{i}$ takes $O (n)$ time. Step (4) updating the REW values of the arcs incident from $τ_{i}$ takes $O (n)$ time. Similarly, step (5) constructing an outgoing arc set with respect to $τ_{i}$ requires $O (n)$ . In steps (6) and (7), the time complexity depends on slack-reclamation methods applied in current schedule and takes maximum $O (n \log n)$ [7–9, 18, 22]. Steps (7) and (8) comparing the values of the arcs in $A_{i}^{'}$ and constructing a new arc set $A_{i}^{''}$ with respect to $τ_{i}$ require $O (n)$ time. Finally, in step (9), searching an arc $a (τ_{i}, τ_{x})$ in $A_{i}^{''}$ with the maximum value of $R E W (τ_{i}, τ_{x})$ requires $O (n)$ time, and step 10 takes constant time to switch the processor mode. Therefore, the time complexity of the online part of LL-PCP is $O (n \log n)$ .

6. Conclusions

This paper reduces priority inversion of real-time task synchronization with speed/voltage switching overhead consideration. The objective is to minimize the priority inversion of a given task set and reduce the leakage energy, provided that the schedulability of tasks is guaranteed. By traversing the vertex in WDG, the scheduling decisions can be done efficiently during the runtime.

For further study, we shall explore the minimization issues of energy consumption on dynamic priority assignment, for example, the Earliest Deadline First Scheduling with Stack Resource Policy [28]. Future research and experiments in these areas may benefit several mobile system designs.

Footnotes

Conflict of Interests

The authors declare that there is no conflict of interests with any financial organization regarding the experimental process and discussion in the paper.

Acknowledgment

The authors would like to thank the National Science Council of the Republic of China, Taiwan, for financially supporting this research under NSC 102-2221-E-025-003, NSC 102-2622-E-025-001-CC3, and NSC-101-2622-E-025-002-CC3.

References

Duarte

Vijaykrishnan

Irwin

M. J.

Tsai

Y. F.

Impact of technology scaling and packaging on dynamic voltage scaling techniques

Proceedings of the 15th Annual IEEE International ASIC/SOC Conference

September 2002

Chen

J.-J.

Kuo

T.-W.

Procrastination for leakage-aware rate-monotonic scheduling on a dynamic voltage scaling processor

Proceedings of the ACM SIGPLAN/SIGBED Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES '06)

June 2006

Ottawa, Canada

153 162

2-s2.0-33746077098

Gruian

Hard real-time scheduling for low-energy using stochastic data and DVS processors

Proceedings of the International Symposium on Low Electronics and Design (ISLPED '01)

August 2001

Huntington Beach, Calif, USA

46 51

2-s2.0-0034878371

Irani

Shukla

Gupta

Algorithms for power savings

Proceedings of the 14th Annual ACM-SIAM Symposium on Discrete Algorithms

2003

Baltimore, Md, USA

37 46

2-s2.0-0038416699

Jejurikar

Gupta

Procrastination scheduling in fixed priority real-time systems

Proceedings of the ACM SIGPLAN/SIGBED Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES '04)

June 2004

57 65

2-s2.0-4544259605

Jejurikar

Pereira

Gupta

Leakage aware dynamic voltage scaling for real-time embedded systems

Proceedings of the 41st Design Automation Conference

June 2004

275 280

2-s2.0-4444368993

Kim

Min

S. L.

A dynamic voltage scaling algorithm for dynamic-priority hard real-time systems using slack time analysis

Proceedings of the Design Automation and Test in Europe (DATE '02)

August 2002

788 797

Kim

Min

S. L.

Dynamic voltage scaling algorithm for fixed-priority real-time systems using work-demand analysis

Proceedings of the 2003 International Symposium on Low Power Electronics and Design (ISLPED '03)

August 2003

New York, NY, USA

ACM Press

396 401

2-s2.0-1542329187

Kim

Min

S. L.

Preemption-aware dynamic voltage scaling in hard real-time systems

Proceedings of the 2004 International Symposium on Lower Power Electronics and Design (ISLPED '04)

August 2004

393 398

2-s2.0-16244372710

10.

Liang

W. Y.

Lai

P. T.

Chiou

C. W.

An energy conservation DVFS algorithm for the android operating system

Journal of Convergence 2010 1 1 93 100

11.

Mochocki

X. S.

Quan

Transition-overhead-aware voltage scheduling for fixed-priority real-time systems

ACM Transactions on Design Automation of Electronic Systems 2007 12 2

2-s2.0-34248328748

10.1145/1230800.1230803

1230803

12.

Pillai

Shin

K. G.

Real-time dynamic voltage scaling for low power embedded operating systems

Proceedings of the 18th ACM Symposium on Operating Systems Principles (SOSP '01)

2001

New York, NY, USA

ACM Press

89 102

13.

Quan

Niu

X. S.

Mochocki

Fixed priority scheduling for reducing overall energy on variable voltage processors

Proceedings of the 25th IEEE International Real-Time Systems Symposium (RTSS '04)

December 2004

309 318

2-s2.0-21644465471

14.

Lee

Y. H.

Reddy

K. P.

Krishna

C. M.

Scheduling techniques for reducing leakage power in hard real-time systems

Proceedings of the 15th Euromicro Conference on Real-Time Systems (ECRTS' 03)

2003

105 112

15.

Jejurikar

Gupta

Dynamic slack reclamation with procrastination scheduling in real-time embedded systems

Proceedings of the 42nd Design Automation Conference (DAC '05)

June 2005

New York, NY, USA

111 116

2-s2.0-27944494362

16.

Niu

Quan

Reducing both dynamic and energy consumption for hard real-time systems

Proceedings of the International Conference on Compilers, Architecture, and Synthesis for Embedded Systems

September 2004

140 148

2-s2.0-29144443691

17.

Baptiste

Scheduling unit tasks to minimize the number of idle periods: a polynomial time algorithm for offline dynamic power management

Proceedings of the 17th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA '06)

January 2006

364 367

2-s2.0-33244494813

10.1145/1109557.1109598

18.

Jejurikar

Gupta

Dynamic voltage scaling for systemwide energy minimization in real-time embedded systems

Proceedings of the International Symposium on Lower Power Electronics and Design (ISLPED '04)

August 2004

78 81

2-s2.0-16244399186

19.

International Technology Roadmap for Semiconductors

2002, http://public.itrs.net

20.

Borkar

Design challenges of technology scaling

IEEE Micro 1999 19 4 23 29

2-s2.0-0032592096

10.1109/40.782564

21.

Huang

Fan

Quan

On-line leakage-aware energy minimization scheduling for hard real-time systems

Proceedings of the 17th Asia and South Pacific Design Automation Conference (ASP-DAC '12)

February 2012

Sydney, Australia

677 682

2-s2.0-84859979316

10.1109/ASPDAC.2012.6165042

22.

Aydin

Melhem

Mossé

Mejia-Alvarez

Power-aware scheduling for periodic real-time tasks

IEEE Transactions on Computers 2004 53 5 584 600

2-s2.0-3042585906

10.1109/TC.2004.1275298

23.

Sha

Rajkumar

Lehoczky

J. P.

Priority inheritance protocols: an approach to real-time synchronization

IEEE Transactions on Computers 1990 39 9 1175 1185

2-s2.0-0025488794

10.1109/12.57058

24.

Silberschatz

A. P.

Galvin

Gagne

Operating System Concepts 2011

John Willey and Sons,

25.

Zhang

Chanson

S. T.

Processor voltage scheduling for real-time tasks with non-preemptible sections

Proceedings of the 23rd Proceedings IEEE Real-Time Systems Symposium (RTSS '02)

December 2002

Austin, Tex, USA

235 245

26.

Jejurikar

Gupta

Energy-aware task scheduling with task synchronization for embedded real-time systems

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 2006 25 6 1024 1036

2-s2.0-33646746385

10.1109/TCAD.2005.855964

27.

Butts

J. A.

Sohi

G. S.

Static power model for architects

Proceedings of the 33rd Annual IEEE/ACM International Symposium on Microarchitecture

December 2000

Monterey, Calif, USA

191 201

2-s2.0-0034427485

28.

Liu

J. W. S.

Real-Time Systems 2000

Upper Saddle River, NJ, USA

Prentice Hall

A Blocking-Aware Scheduling for Real-Time Task Synchronization Using a Leakage-Controlled Method

Abstract

1. Introduction

2. System Model

2.1. Task Model

2.2. Power Model

2.3. Resource Access Model

3. Motivating Example

4. Latency Locking PCP

Step 1.

Step 2.

Step 3.

Algorithm 1: LL-PCP algorithm.

Definition 1 (expected sleeping interval, ESI).

Definition 2 (reduction of priority inversion time, RPI).

Definition 3.

5. Experimental Results

6. Conclusions

Footnotes

Conflict of Interests

Acknowledgment

References