An Intelligent Power Save Mode Mechanism for IEEE ... - IEEE Xplore

1 downloads 0 Views 316KB Size Report
mobile device wireless interface. IEEE 802.11 wireless local area network (WLAN) power save mode (PSM) minimizes the mobile device energy consumption ...
Globecom 2012 - Symposium on Selected Areas in Communications

An Intelligent Power Save Mode Mechanism for IEEE 802.11 WLAN Haleh Tabrizi† , Golnaz Farhadi⋆ , John Cioffi†∗ †

Department of Electrical Engineering, Stanford University ⋆ Fujitsu Laboratories of America, Inc. ∗ Department of Computer Science, King Abdulaziz University Emails: {htabrizi, cioffi}@stanford.edu, and [email protected] Abstract—Maximizing the run-time of battery-limited mobile devices is contingent on efficient energy management at the mobile device wireless interface. IEEE 802.11 wireless local area network (WLAN) power save mode (PSM) minimizes the mobile device energy consumption by allowing the device to de-activate its wireless interface periodically while the access point (AP) stores its incoming packets. Further increase in energy conservation occurs through the proposed decision-making algorithm, Intelligent-PSM (IPSM), that determines dynamic, as opposed to fixed, activation times based on the state of the mobile user. IPSM can be implemented with minimal change to the standard PSM transmission scheme. Simulation results show that with a maximum allowable packet delay of 1 second and limited AP buffering capacity, energy conservation of up to about 37% compared to standard PSM can be achieved.

I. I NTRODUCTION The increasing demand for mobile computing has led to a large increase in the deployment of WiFi APs and the ever-increasing usage of mobile devices with limited energy capacity. The general desire is to be able to interact with information anywhere and at anytime. However, the finite battery capacity of mobile devices limits their run-time. Wireless cards have shown to account for about 50% of the total energy consumption in current handheld devices [1]. Therefore, energy conservation at the wireless interface is a critical requirement to elongate the portable device run-time. According to the 802.11 standard, all stations in PSM (denoted by STA) that are associated with an AP are synchronized to activate their wireless interface at the same time. At this time, the AP sends a beacon frame that indicates the packets it has buffered. If the STA has a buffered packet, it sends a PS-Poll frame to the AP and the AP sends the corresponding buffered packets to the STA. Otherwise, the STA will deactivate and wait until the next beacon time. One main issue with this protocol is that the STA has to activate at every beacon time even when there are no buffered packets. The energy consumed by the mobile device to activate from an off state is comparable to the energy consumed when receiving a packet. Hence if the number of activations can be reduced without introducing long delays, the energy consumption can be reduced significantly. Much work has been dedicated to modifying PSM to reduce mobile station energy consumption further. One of the earlier papers, [2], shows that PSM can save up to 90% of the wireless ⋆

This work was done while G. Farhadi was with Stanford University.

978-1-4673-0921-9/12/$31.00 ©2012 IEEE

interface energy consumption at the expense of long delays. To confront the long delay problem, it proposes the Bounded Slowdown Protocol (BSP) that adaptively slows down the activation frequency of the STA during idle times. A cross layer energy manager (XEM) that dynamically changes its energy conserving strategy based on the user application and network traffic is proposed in [1]. XEM switches the wireless interface between PSM and off mode. A stochastic timer-based power management model is developed in [3]. Based on this model, the probabilities that the STA is in active, doze, or idle mode are derived. An algorithm that can optimize the STA idle and doze durations is then obtained. A scheduled power saving mode (SPSM) algorithm based on time slicing is proposed in [4]. It schedules APs to deliver pending data at designated time slices and adjusts the power state of mobile STAs adaptively. An opportunistic PSM (OPSM) algorithm is proposed in [5]. The STA waits to activate at the time that requires the least amount of energy, i.e, the time that the AP is not serving any other STA. [6] proposes a scheme for congestion scenarios when all buffered packets of a STA cannot be transmitted in one beacon interval. The proposed solution is to mark a group of stations that should activate at each beacon interval. Other than modifying the standard PSM, other energy conserving strategies have also been analyzed. Bluetooth consumes less energy than WiFi, hence [7] examines switching between WiFi and Bluetooth interfaces based on the transmission range and application bandwidth requirements. This paper explores how dynamic selection of STA activation times in infrastructure Basic Service Set (BSS) can increase its run-time. The objective is to minimize energy consumption with minimal change in the standard PSM scheme. The only time the AP can communicate with the STA is when the STA activates and receives a beacon. At this time, the AP predicts the optimal next beacon time that the STA should activate and communicates that time through its beacon. The selected activation times are integer multiples of beacon arrival times. Hence, the existing AP beacon transmission method does not need to be modified. Only 3 to 4 bits representing an integer multiple of beacon intervals is added to the beacon frame. This problem is modeled as a Markov decision process (MDP) and a cost function associated with each state-action pair is defined. Value iteration is then used to find an optimal activation time policy.

3460

Fig. 2. Fig. 1.

MMPP state diagram and transition probability matrix P .

Access point buffer and associated station.

II. S YSTEM M ODEL The system model considered here consists of a single AP and a single associated station as shown in Figure 1. A similar analysis follows when there are more than one associated stations. The WiFi AP is modeled as a buffer with finite capacity of q IP packets of size B bytes. More than q buffered packets are discarded because the AP does not have enough storage capacity or the packets expire. Packets aimed at each associated station arrive at the AP with some known distribution. The AP transmits beacon frames at every beacon interval b. Time is slotted with slot duration equal to one beacon interval. The time at the beginning of each slot (or beacon interval) is represented by t1 , t2 , ... The beacon contains the STA’s next activation time, which is equivalent to the number of beacon intervals that the STA can sleep. The STA activation time, an integer multiple of b, can be conveyed to the STA through some additional bits in the beacon frame. This is the only required addition to the traditional PSM communication model. Furthermore, at every tk , a cost amount is assigned to the AP. The cost is equivalent to the STA average energy required to receive the buffered packets and a cost proportional to the probability of packet drops until the STA’s next activation time. If the STA is asleep for too long such that the STA packets at the AP overflow the q capacity, they will be dropped. The goal is to determine STA activation times in such a manner that minimizes the long term cost. The activation decision at each decision epoch tk is based on the current system state and the statistics of packet arrivals. It is assumed that the AP is capable of learning the network and application statistics. According to 802.11 PSM, the STA can switch between three different modes of operation: 1) sleep mode, during which the STA cannot receive or transmit packets and consumes the least amount of energy, 2) active mode, during which the STA can transmit and receive packets, and 3) transition mode, during which the STA transitions from sleep to active mode. The amount of energy consumed when transitioning from active to sleep mode is negligible and hence omitted in this analysis. The amount of power consumed at each state is denoted by Psleep , Pactive , and Ptrans , respectively. Let Tpkt , Tb , and Ttrans be the time required by the STA to receive a single packet of size B, receive a beacon frame, and the time required for the STA to switch from sleep mode to active mode, respectively. Tpkt is the average time required for receiving a packet, which includes the STA back-off time,

wait times, and transmission of PS-Poll frame and receiving acknowledgment frames. As is apparent from definitions, Tpkt and Tb depend on the downlink data rate of the AP. Although both downlink and uplink traffic occupy the same spectrum, only downlink data traffic is considered here. The packet arrival model is considered to be a Markov Modulated Poisson Process (MMPP). An MMPP is a stochastic arrival process that differs from a Poisson process in that the arrival rate is not constant. The instantaneous arrival rate is given by the state of a Markov process. For simplicity, a twostate MMPP is considered here. The system is either in an off state or an on state as depicted in Figure 2. On average, the off state represents idle times and the on state represents packet bursts. If in the on state, packets arrive according to a Poisson process of rate λ, and if in the off state, there are no packet arrivals. The transition from one state to another occurs only at the beacon interval times. The MMPP is described by a state transition probability matrix P indicated in Figure 2. The transition probability from the on state to the off state is α and from off state to on state is β. Let L = {0, 1, ..., q} be the set of possible number of packets buffered at the AP. The function e(l) : L → R+ denotes the amount of energy consumed by a STA for receiving l ∈ L packets from the AP. After the STA receives all the buffered packets l, it goes back to sleep until the next scheduled activation time. It is assumed that the packets destined to the STA that arrive at the AP during the transmission of the l packets are not transmitted instantaneously, but are buffered and constitute the next buffer state. When the STA activates at the beginning of a beacon interval, the amount of energy it consumes in one slot duration is equal to the sum of four different sources of energy consumption: 1) the energy consumed for one sleep to active transition, 2) energy required for receiving one beacon frame, 3) energy required for receiving l packets, and 4) energy consumed when the STA is sleep for the remaining duration of the beacon interval. Mathematically,

e(l) = Ttrans Ptrans + Tb Pactive + lTpkt Pactive

(1)

+(b − Ttrans − Tb − lTpkt )Psleep . It is assumed that the time required to receive the maximum number of packets that can be buffered at the AP is less than a beacon interval, i.e., b ≥ Ttrans + Tb + qTpkt .

3461

III. IPSM: P ROBLEM F ORMULATION A. States and Actions The problem of dynamically selecting STA activation times can be modeled as a discounted MDP. Three elements determine the state x = {l, m, a ˆ} of the system: 1) l ∈ {0, 1, ..., q}, the number of packets in the buffer, 2) m ∈ {on, of f }, the MMPP state, and 3) a ˆ ∈ {1, 2, ..., W }, a timer that determines the next activation time of the STA. The parameter W determines the maximum number of beacon intervals that a packet can be delayed in reaching its destination, and it is specified based on the underlying application delay requirements. The number of possible states at each decision epoch is thus equal to 2 · W · (q + 1), and the state space X is given by, X = {0, 1, ..., q} × {on, of f } × {1, ..., W } .

(2)

The control action is denoted by a and determines the next beacon time that the STA should activate. The AP computes a(x) based on the state of the system and communicates the information to the STA through the current beacon. The STA listens to the beacon, and by sending PS-Poll frame to the AP, it obtains as many of its buffered packets as possible. When finished receiving all the buffered packets, the STA goes to sleep and activates the next a beacon intervals. Let A denote the set of all possible actions: A = {a|a ∈ {1, 2, ..., W, φ}} ,

(3)

where φ indicates no decision. At time tk with decision ak the STA will be asleep for ak beacon intervals and hence will skip the next ak − 1 beacon frames. No communication can occur between the STA and AP during that interval and hence no decision will be made for the states when the STA is asleep. The time epochs tk′ at which no decision is made, the action ak′ is identified as φ. B. Transition Probabilities The next instant that an activation decision can be communicated to the STA is its next activation time, that is a beacon intervals from the previous decision time. Figure 3 shows beacons that are transmitted periodically at equal distances b apart. In this figure, at time t1 the first beacon carrying a decision a1 = 2 is transmitted. This decision indicates that the STA can sleep during the next beacon and activate in two beacon intervals. Furthermore, this decision indicates that the next decision will be made in two beacon intervals from the first decision epoch and hence a2 = φ. At time t3 the decision is a3 = 3 and indicates that the STA should activate in three beacon intervals and the third decision a6 will be communicated at that time. Figure 4 represents this decision process as an MDP. The figure corresponds to the same actions taken in Figure 3. This process begins at state x1 ∈ X, and according to this state, some action a1 ∈ A is selected. As a result of this action, the state of the MDP randomly transitions to some successor state x2 drawn according to x2 ∼ Px1 ,a1 . Then at state x2 , another action a2 is selected and the state transitions again according

Fig. 3.

STA activates only at certain beacon times.

to x3 ∼ Px2 ,a2 and so forth. Let xk be the state of the system at time tk . If the system is in state xk , the next state xk+1 is determined by the MMPP state mk and the corresponding packet arrival rate. Hence, given state xk ∈ X and a control action ak ∈ A, the next state xk+1 is given by a stochastic function that will be defined next. Given current state xk = (lk , mk , a ˆk ) and action ak , the timer a ˆk+1 is deterministic and is determined as follows:

a ˆk+1 =



a ˆk − 1 ak

if ak = φ otherwise

(4)

The only time the AP can select an action (i.e. a 6= φ) is when a ˆk = 1. At each epoch that no action is selected (ak = φ), a ˆk+1 is reduced by one until it reaches a state when a ˆk = 1 and that’s when action ak 6= φ is selected and a ˆk+1 resets to the new value of ak , i.e. a ˆk+1 = ak . When MMPP state m = on, packets arrive at the AP according to a Poisson process of rate λ during the beacon interval. Let P oiss(λ, d) denote the probability of d packet arrivals in one beacon interval with average arrival rate of λ. When a ˆk = 1, that is the STA activates, it will receive all the buffered packets lk and hence the number of buffered packets at the next time epoch lk+1 will be independent of lk . However, if the STA does not activate (ˆ ak 6= 1), the next state packets lk+1 will be equal to the number of arrivals plus the previous number of packets in buffer. Denote the remaining number of packets after taking action ak with rk = lk 1{ˆ ak 6= 1}. The number of arrivals in the kth beacon interval denoted by dk is: dk = lk+1 − rk ,

(5)

With the definition of a ˆ in equation (4) and the above mentioned definitions of dk and P oiss(.), the probability of transitioning from state (l, m, a ˆ) to (l′ , m′ , .) is given by the equations below:

P(l,m,ˆa),(l′ ,m′ ,.)

 (1 − α)P oiss(λ, d)    αP oiss(λ, d) = (1 − β)1{d = 0}    β1{d = 0}

if if if if

(on, on) (on, of f ) (of f, of f ) (of f, on)

where values inside the parenthesis after the ‘if’ statements correspond to the current and next MMPP states: (m, m′ ). When the current MMPP state m = of f , the number of arrivals d has to be zero and hence the indicator function 1{.}.

3462

(6)

TABLE I S YSTEM PARAMETERS Parameter b γ bit-rate Tpkt Tb Ttrans Fig. 4.

Value 100 0.98 11Mb/s 1.5 1 0.8

Parameter q W B Pactive Psleep Ptrans

Value 100 10 512 Bytes 1.4 0.045 2.3

Markov decision process.

J(xk )

IV. IPSM: M ETHOD OF S OLUTION A. Cost Function Associated with each state xk , for k = 1, 2, ... is a cost function C(xk ) that is defined as follows: C(xk ) = E(xk ) + D(xk ),

E(xk ) = e(lk )1{ˆ ak = 1} + bPsleep 1{ˆ ak 6= 1},

(8)

If a ˆk = 1, the STA is activating and the buffered packets are transmitted to the STA. The energy consumed by the STA is thus equal to the energy required to receive lk packets, e(lk ). Otherwise, the energy consumption is equal to the energy required for the STA to sleep for one beacon interval, bPsleep . If the total number of packets at the buffer is greater than the buffer capacity q, the excess packets will be dropped. A quadratic drop cost function is considered, which is added to the total cost at each epoch. The drop cost at each state is:

min [ C(xk )

+

γ

+

(7)

where E(xk ) is the energy consumed by the STA in one beacon interval at state xk and D(xk ) is the cost of dropped packets. The relationship between the function E(.) and the energy consumption function e(.) defined in equation (1) is described by:

=

a∈A q−r Xk

γ

(11)

Pxk ,xk+1 · J(rk + d, mk+1 , a ˆk+1 )

d=0 ∞ X

Pxk ,xk+1 · J(q, mk+1 , a ˆk+1 ) ]

d=q−rk +1

If the number of packet arrivals d does not overload the buffer (d ≤ q −rk ), the next buffer state is equal to the number of arrivals plus the remainder of packets from the previous state, i.e, lk+1 = rk +d. If the number of arrivals does overload the buffer, the excess packets will be dropped and the next state will correspond to a full buffer, lk+1 = q. There are various algorithms for solving MDPs. For an infinite horizon MDP with finite state and action spaces, value iteration is an efficient algorithm and it is the method employed here to obtain an optimal decision policy. V. MDP S OLUTION AND R ESULTS

Table I contains the values of the parameters used in obtaining the optimum policies and performing simulations. Time parameters are in milliseconds and power parameters are in Watts. These parameters depend on the specific WiFi card used. Values similar to [8] have been selected here. The parameters that determine the system model are the MMPP probabilities α and β, and the packet arrival rate λ. Three  ∞ P  η 2 parameters 1) q, the maximum number of packets that can d P oiss(λ, d) if mk = on (9) be buffered at the AP, 2) W , the maximum allowable packet D(xk ) = d=q−rk +1  0 if mk = of f delay, and 3) η, the weight of the packet drop cost, are determined by the user application and selected by the AP. where η is a normalization factor. If the current MMPP state Simulations have been performed for different system model is off, there are no packet arrivals and hence the drop cost is parameters (as indicated below) and the variables are set to zero. Otherwise, the drop cost depends on the probability of q = 100 and η = 0.3. A maximum delay of 1 second the number of packet arrivals in one beacon interval. corresponding to W = 10 (beacon intervals) has been selected. B. Dynamic Programming Solution A. Optimum Policy In order to solve the dynamic programming problem, the Optimum policies for three different combinations of (α, cost-to-go function at state x given by J(x) is written in the β) values and λ varying between 5 and 30 packet arrivals form of Bellman equation: per beacon interval, b, have been obtained. The resulting X policies are shown in table II. As explained previously, (α, Pxx′ (a)J(x′ ))] (10) β) and λ determine a system model and it is assumed that the J(x) = min[C(x, a) + γ a∈A x′ AP knows these values for its associated STA. For example, where γ is the discount factor, and Pxx′ (a) is the probability assume (α, β) = (0.7, 0.3), and λ = 20 pkts/b. According of transitioning to state x′ when action a is taken at state x to the optimum policy in table II, when the system state and it is obtained using equations (4) and (6). Equation (10) xk = (lk , on, 1), the optimum action ak = 7 and when can be simplified and rewritten as follows: xk = (lk , off, 1), the optimum action ak = 9.

3463

TABLE II IPSM P OLICY

(α, β) = (0.3, 0.7) (α, β) = (0.5, 0.5) (α, β) = (0.7, 0.3)

5 10 10 10 10 10 10

λ (Pkts/b) 10 15 20 10 6 4 10 7 5 10 7 4 10 9 6 10 10 7 10 10 9

30 2 3 2 3 3 4

This trend can be seen in table II and in the simulation results in Figure 5. For a specific value of (α, β), as λ increases, the energy consumption of the IPSM algorithm becomes closer to the PSM energy consumption. For a fixed λ, the larger the value of α and the smaller the value of β, the higher the probability that the system is in the off state, which means no packet arrivals. When there are no packet arrivals, the STA can skip more beacons and consume less energy compared to PSM. This trend can be viewed in the policies obtained in table II and the simulation results in Figure 5. Comparison of the policies in table II for (α, β) = (0.7, 0.3) with the other two sets, shows that the optimum actions have larger values. Simulation results in Figure 5 also confirm the higher gains obtained for higher α and lower β values.

On Off On Off On Off

Energy saving comparison of IPSM and PSM 38 36

IPSM Efficiency (%)

34 32

VI. S UMMARY

30

β = [0.3, 0.7] β = [0.5, 0.5] β = [0.7, 0.3]

28 26 24 22 5

10

15

20

25

30

λ

Fig. 5. Comparison of IPSM and standard PSM with q = 100 and W = 10

In simple words, at time epoch tk , the AP observes the state of the system xk . If a ˆk 6= 1, it knows that the STA is not going to activate at this beacon time, so the selected action ak = φ. It subtracts one from a ˆk for the next state: a ˆk+1 = a ˆk − 1. If the observed state a ˆk = 1, then the AP knows that the STA is scheduled to activate at this beacon time. AP observes mk , and according to table II it selects action ak and includes it in the beacon frame that it broadcasts. After receiving the beacon, the STA obtains its buffered packets and goes to sleep for the next ak beacon intervals, i.e. a maximum delay of ak b msec. The AP sets a ˆk+1 = ak for the next state and subtracts 1 every beacon time after that to keep track of the next activation time. B. Simulation Results Simulations are performed for the system model parameters indicated in tables I and II. The average amount of energy consumed incorporating the proposed IPSM algorithm is compared to the average energy consumed using the standard PSM model. The corresponding ratio of the average energies is shown in Figure 5. The horizontal axis represents packet arrival rate λ and the vertical axis shows the energy consumption ratio of proposed IPSM to standard PSM. The energy in each case is averaged over 100 decision epochs. In all scenarios, the number of dropped packets is zero. In general, the larger the action value, the larger the energy conservation is. The larger the value of λ, the closer the performance of the IPSM algorithm is to the performance of standard PSM. To prevent buffer overflow and hence packet loss, as the number of packets arriving at the AP increases, the AP selects a smaller sleep time to empty its buffer more frequently. For a specific set of (α, β), as λ increases, the optimum action decreases.

This paper examines dynamic activation-time selection for stations in PSM as opposed to fixed activation times. The objective is to minimize power consumption at the mobile station by reducing the number of its activation times. Based on the maximum allowable packet delay and AP buffer capacity, an optimal activation time policy is obtained through dynamic programming at the AP. The obtained policy is stored at the AP and according to the packet arrival rates, the next STA activation time is conveyed to the STA through beacons. It is observed that this approach is most effective when the packet arrival rate is low and the packet idle durations are long. In this situation, the proposed IPSM scheme obtains an energy efficiency of up to about 37% compared to standard PSM. If considering multiple associated STAs, all STAs’ transmission models can be incorporated into the general system model. Each STA’s activation time is then selected dynamically to further eliminate STA idle times that are a main source of energy inefficiency. R EFERENCES [1] G. Anastasi, M. Conti, E. Gregori, and A. Passarella, “802.11 powersaving mode for mobile computing in wi-fi hotspots: Limitations, enhancements and open issues,” Wireless Networks, vol. 14, no. 6, 2008. [2] R. Krashinsky and H. Balakrishnan, “Minimizing energy for wireless web access with bounded slowdown,” Wireless Networks, vol. 11, no. 1-2, pp. 135–148, 2005. [3] Y. hua Zhu and V. C. Leung, “Efficient power management for infrastructure ieee 802.11 wlans,” IEEE Trans. on Wireless Commun., vol. 9, no. 7, pp. 2196–2205, 2010. [4] Y. He, R. Yuan, X. Ma, J. Li, and C. Wang, “Scheduled psm for minimizing energy in wireless lans,” IEEE Int. Conf. on Network Protocols (ICNP), pp. 154–163, 2007. [5] P. Agrawal, A. Kumar, J. Kuri, M. Panda, V. Navda, and R. Ramjee, “Opsm - opportunistic power save mode for infrastructure ieee 802.11 wlan,” IEEE Int. Conf. on Commun. Workshops (ICC), pp. 1–6, 2010. [6] J.-R. Lee and D.-H. Cho, “An energy-efficient downlink multiple access control considering congestion in wireless lans,” IEEE Communication Letters, vol. 10, pp. 405–407, 2006. [7] T. Pering, Y. Agarwal, R. Gupta, and R. Want, “Coolspots: reducing the power consumption of wireless mobile devices with multiple radio interfaces,” Proceedings of the 4th Int. Conf. on Mobile systems, applications and services (MobiSys), pp. 220–232, 2006. [8] E.-S. Jung and N. Vaidya, “An energy efficient mac protocol for wireless lans,” Proceedings of the Twenty-First Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), vol. 3, pp. 1756–1764, 2002.

3464