Adaptive Traffic-Aware PSM Mechanism for IEEE 802.11 WLANs

2.1 PSM in IEEE 802.11 WLANs

In a WLAN under an infrastructure mode, stations connect to the Internet through the AP where the downlink traffic from the AP to the stations is a major part. A station using the PSM can sleep with low power when it does not receive or transmit any data. Accordingly, the AP buffers the frames for sleeping stations and periodically announces the buffering status to all associated stations. The buffering status is recorded in the field of the traffic indication map (TIM) in beacon frames. The PSM-enabled stations can wake up periodically to listen to beacon frames, and reply to the AP by sending power save polls (PS-Poll) frames if their data frames are buffered. Then, the AP sends a data frame to the station whose PS-Poll frame wins the channel competition. This station repeats sending PS-Poll frames to fetch its buffered data frames one by one. If there is no more buffered data, the station will fall asleep and wait for the next wake-up epoch. This process is shown in Figure 1.

Figure 1

Packet Exchanges in the Standard PSM.

In the standard PSM, the AP stores the TIM in beacon frames and periodically broadcasts them. The AP’s period is called the beacon interval (BI). Accordingly, the stations periodically wake up to listen to beacon frames. The listening period is called the listen interval (LI), which must be an integer multiple of BI, i.e., LI = n * BI, n = 1, 2, 3, …. Different stations can set different values of LI. For example, in Figure 1, the LI of Station1 is 1 * BI, while the LI of Station2 is 2 * BI. The default setting of BI = 100 ms and LI = BI cannot adapt to various wireless applications because the changes of traffic and channel conditions are not considered. When the traffic is heavy, the AP will buffer much data or even overflow, and the delay of frames will increase greatly. When the traffic is light, the stations will waste a lot of energy on frequently wake-ups.

2.2 Energy-Saving Schemes for Wireless Clients

The basic idea of energy saving for mobile devices is adjusting the working manners, including the usage of advanced hardware [7, 24], the improvement of network protocols, and the assistance of a third party. For example, Kumar and Lu [18] and Peng et al. [23] found that cloud assistance is practicable for energy saving and performance extension by offloading computation from wireless clients to cloud. In this section, we focus on discussing the improvement of the PSM in the standard of WLANs, IEEE 802.11.

Some researchers slightly adjusted the sleep/wake-up mechanism of the PSM to improve communication performance. Zhu [32] allowed a PSM-enabled station fall asleep after receiving two continuous beacon frames when the AP did not buffer data for it. To prevent channel collisions, Lin et al. [19] designed a new wake-up mechanism where a beacon frame only informs at most one station to fetch its buffered data. Moreover, He et al. [12] changed the structure of the TIM, which indicates the time slices of communication for each station within one BI. Zafer and Eytan [31] presented a rate-control approach to minimize the transmission energy expenditure while serving the data with the deadline constraints.

Some advanced PSM strategies are designed for higher levels, for example, referring to the design of transport protocols [2] or the adaption to some application traffic. For example, Krashinsky and Balakrishnan [17] proposed a power-saving strategy for web applications with TCP traffic. Mur et al. [21] proposed an adaptive mechanism according to the traffic characters in specific applications. Chandra and Vahdat [3] proposed a server side traffic shaping mechanism that saves much energy for all the stream formats without any data loss. In addition, Chen and Jin [5] simplified the network environment of mobility and designed an improved PSM strategy based on the location prediction of stations.

Our previous work [6] found some insights into improving the performance of the IEEE 802.11 PSM. In this article, we propose the APSM to improve the energy efficiency of stations.¹ The most prominent characteristic of the APSM is that it considers the channel status and the traffic load of all associated stations. The AP ranks the stations according to the traffic load and helps the stations select the optimal sleeping schemes to save more energy. Second, the APSM is easily compatible with IEEE 802.11, as it makes the best of the existing parameters in the standard process of data exchanging. Third, unlike the many previous attempts of PSM improvement [11, 27, 29], the APSM has been implemented and evaluated in the popular network simulator NS-2 [28], instead of some customized simulators.

3.1 Network Model

In this article, we consider a WLAN that consists of one AP and several stations. Let c denote the number of stations associated with the AP. The total number of buffered frames in the AP is denoted as Q, which is expressed in Eq. (1), where q_i is the length of data frames with the destination of the i^th associated station. The AP can evaluate the busy status of the wireless channel (denoted as ω) by Q, and the traffic status for each station (denoted as τ_i) by q_i, i = 1, …, c.

In our model, the AP has an important role for two reasons. The first is the core status of the AP during communication. When some stations join/leave, they inform the AP by association or disassociation frames. The AP broadcasts beacon frames to all associated stations, while the stations always receive or transmit packets through the AP. Therefore, the AP can detect the concrete traffic amount and the topology changes of the network. The second is that the AP has an external power supply with a powerful computing capability. Furthermore, the AP can buffer data for sleeping stations and deploy the main algorithms of energy-saving strategies.

Therefore, the design of the APSM shall meet the following principles, which require the cooperation of the AP and stations:

• The AP shall play a leading role in the APSM. All data destined to the stations using the APSM must be buffered in the AP until the corresponding PS-Poll frames arrive. The AP shall follow some strategies to rank the traffic load for every station, which helps control the operations of data fetching.

• The complicated algorithms of the APSM are mainly deployed in the side of the AP. The AP can make the best of its powerful capability and its awareness of the changes of channel and traffic.

• Beacon frames shall include sufficient network information with low cost. The associated station is aware of the changes of channel and traffic, and then selects an optimal energy-saving strategy accordingly.

3.2 Improved Strategies in AP

In the APSM, the AP has a smart buffering scheme and provides the information about channel and traffic for all stations. It shall achieve the following functions:

1. Record the channel status and the number of associated stations. In a BSS, the AP utilizes the management frames to obtain the information of the channel and associated stations, such as probe request/response, authentication/deauthentication, association request/response, and disassociation.

2. Be aware of the traffic load based on the buffer statuses. The AP easily knows the number of data frames destined to each station and estimates the traffic rate within the BI accordingly.

3. Maintain the fairness among stations when fetching buffered data. The traffic load and characteristics are different in various wireless applications. Each station adapts its frequency of data fetching according to the traffic load and its prioritization. A heavy-load station can gain more communication chances and a light-load station may not starve.

4. Take full advantages of beacon frames to periodically broadcast the information of the network. Then, the stations will select different energy-saving strategies according to the awareness of the network and traffic information.

The APSM follows the general steps of the standard PSM but updates the structure of beacon frame by redesigning some standard fields in Ref. [13], as shown in Figure 2. For example, the traffic indication virtual bitmap that generates a TIM takes up 251 bytes as the association ID (AID) is in [0, 2007]. The i^th bit is 0 if there is no data frame buffered for the station whose AID is i, i∈[0, 2007]. In general applications, one AP simultaneously supports < 100 stations. Therefore, the APSM revises the format of the TIM by constricting the range of AID as [0, 991], where at most 992 (=124 × 8) stations can be supported. Then, the traffic statuses of all stations are following, which are recorded in a new field of Traffic_status with 124 bytes. Similarly, one bit is assigned to present the traffic status of one station according to its AID. Finally, the channel status is assigned to 1 bit, and the number of associated stations is stored in the field of Num_bssid with 15 bits. Obviously, the APSM does not change the basic structure of the beacon frame. This is the main reason why the APSM is compatible with the standard PSM.

Figure 2

The Updated Beacon Frame in APSM.

Next, the AP designs the following rules to solve the fairness problem. The stations with more buffered data shall have high priority to access the AP, while the stations with light traffic load shall be dealt with in an appropriated speed.

• Let the stations with more buffered frames hold higher priority to fetch their data. The corresponding bit of Traffic_status, τ_i, is calculated as Eq. (2), where κ denotes the number of stations that have buffered data in the AP.

  (2)τi = {1ifqiQ ≥ 1κ,0else. (2)

• Estimate the busy level of the channel with the corresponding bit of busy_status, denoted as ω. According to Eq. (3), when at least a half of the stations join in the channel competition or the occupied ratio of buffer is larger than a threshold ε∈(0, 1), the channel will be regarded as busy and ω = 1. Otherwise, ω is set as 0. Here, M is the maximum value of buffer size, and ε = 0.8 in the following:

  (3)ω = {1if Q≥εM||κ≥c2,0else. (3)

In summary, the APSM makes the station that has high traffic rate be awarded with high priority to retrieve buffered data while avoiding the light-traffic stations being chronically hungry.

3.3 Improved Strategies in Stations

As we know, the standard PSM allows stations to set different values of LI for different applications, but fails for the lack of adjustment strategies when network topology or traffic workload change. In the APSM, the stations dynamically adapt to the changes of network conditions to save energy while maintaining a satisfying performance. The stations periodically wake up to receive beacon frames that include the channel status, the changes in the BSS, and the traffic status of each station. Then, the stations can select different wake-up/sleep schemes and fetch buffered data according to the information from beacon frames. Therefore, the stations can obtain high energy efficiency.

The corresponding bit of TIM_i indicates whether the i^th station has buffered data in the AP. When TIM_i =0, the station falls asleep as what it did in the standard PSM. Before that, it is required to update the value of LI and attempt to sleep longer to save energy. Suppose the maximal delay boundary of some application is α. If α <4BI, the station will use the standard PSM by default values. Otherwise, the value of LI is changed as Eq. (4), where LI′i and LI_i denote the updated value and the current value of the i^th station, respectively.

When the i^th station has buffered data in the AP (i.e., TIM_i =1), it will select energy-saving strategies according to Table 1, where δ is a random positive integer, cw_i is the contention window for the i^th station, and t_s is a slot time. The i^th station updates its LI′i according to the value of ω, τ_i, and the current LI_i. When suffering from channel congestions, this station adopts the backoff timer (by the units of slot time) β′ for fetching its buffered data. Obviously, β′ is a function of cw_i, which changes according to the binary exponential backoff algorithm in IEEE 802.11.

4.1 Analysis of Energy Efficiency

In this subsection, the energy-efficiency metric R_ee is mainly used to evaluate the energy efficiency of stations, which records the amount of data transmitted successfully with the energy computation of 1 J. We compare the APSM with the standard PSM using default settings (referred to as PSM, BI = 100 ms, LI_i = BI, and cw_i follows binary exponential backoff), i = 1, 2, …, c. To analyze the effectiveness of the adaptive BI and LI_i, we also introduce an “incomplete version” of the APSM (referred to as Scheme-1), which only adjusts the congestion backoff timer for stations like the APSM. For easy comparison, one performance index of R_ee for comparing the APSM and Scheme-1 against the PSM is introduced, ηee = (Ree − ReeS) / ReeS × 100%. The notations with the superscript “S” refer to the standard PSM, whereas that without superscripts refers to the APSM or Scheme-1. Obviously, a positive value of η_ee indicates the improvement over PSM.

First, we consider the network with three stations as an example (c = 3). Each simulation experiment lasts 20 s, where the size of data frame is fixed by 512 bytes and the average interval of arrival frames are different for stations with λ₁ : λ₂ : λ₃ = 1 : 2 : 3.

Table 3 shows that the APSM outperforms the PSM in improving energy efficiency under different distributions. The APSM achieves the highest R_ee among three schemes when the traffic is not very heavy. Scheme-1 performs much worse than the APSM as it does not adjust BI and LI_i. That is, the benefit of the adaptive BI and LI_i is significant. We also find that the improvement due to the adaptive backoff timer is minor as Scheme-1 only outperforms the PSM a little under most cases. However, an exception occurs under CBR traffic when ρ = 0.6. The PSM is the best scheme because the adaptive actions of the APSM and Scheme-1 have limited effects on heavy constant traffic but waste energy on the frequent unnecessary wake-ups of wireless clients.

The performance indexes η_ee of the APSM and Scheme-1 are positive, and the improvement of the APSM over the PSM is most significant when the traffic is not heavy. As shown in Figure 3, under the EXP distribution of traffic with ρ = 0.1, the APSM improves energy efficiency by 115% compared with the PSM. That is, the APSM, which employs all adaptive parameters, performs best among three schemes. However, the advantage of the APSM decreases with the increase of ρ.

Figure 3

η_ee of the APSM and Scheme-1 Under EXP Traffic, When c = 3.

The adaptive BI and LI_i jointly play major roles in improving energy efficiency as the APSM has a larger η_ee than Scheme-1. For example, as shown in Figure 3, the energy efficiency of Scheme-1 is improved as much as 22%, while the η_ee of the APSM is 51% under the middle level of EXP traffic with ρ = 0.4. At the same time, the differences of η_ee between the APSM and Scheme-1 decrease with ρ. Therefore, the adaptive BIand LI_i are helpful to save energy; however, their influence decreases when the traffic level increases. The traffic under PAR distribution has the similar tendency. However, the tendency of CBR traffic is different when the traffic is heavy (e.g., when ρ≥0.4, as shown in Table 3).

Next, we show that the APSM is also energy efficient in a WLAN with more clients. For simplicity, all clients are assumed to be symmetrical (i.e., λ_i = λ_j, i ≠ j). Two traffic patterns are considered: the total traffic load remains constant, and the total traffic increases with the number of clients.

In Figure 4A, the total traffic workload is fixed as a constant with ρ = 0.3, while the number of stations increases from 4 to 30. We can find that the stations using the APSM mostly have significantly higher energy efficiency than those using Scheme-1 and the PSM when the total traffic workload is light. However, the advantage of Scheme-1 over the PSM is not obvious when the number of stations increases, as the improvement due to the adaptive backoff timer is minor. In Figure 4B, the traffic load becomes heavier when the number of stations increases, where λ_i = 20 ms under CBR traffic. When c ≤ 11, the traffic is not heavy as ρ < 0.6, the energy efficiency of stations in APSM is highest among three schemes. When 11 < c ≤ 16, the traffic is heavy as 0.6 < ρ < 1, the advantage of the APSM in energy efficiency becomes minor or even disappears. Especially, Scheme-1 is worse than the PSM under middle/heavy constant traffic because the adaptive backoff timer wastes energy on many unnecessary wake-ups of clients. The traffic under EXP and PAR distributions has the similar tendency.

Figure 4

R_ee Under CBR Traffic vs. c.

(A) [ρ = 0.3]; (B) [λ_i = 20 ms].

4.2 Performance of the APSM

Our previous work [6] has shown that the standard PSM extends the delay within an acceptable range. Next, we analyze the delay of traffic in stations using the APSM.

Similar to Subsection 4.1, we first use a network with three stations as an example. The PAR traffic is considered when ρ increases from 0.1 to 0.6 and λ₁ : λ₂ : λ₃ = 1 : 2 : 3. As shown in Figure 5A, the data delay in the APSM is slightly larger than that in the PSM when ρ is <0.45. At the same time, Scheme-1 and the PSM have a similar delay. When the traffic is >0.5, the delay of the PSM increases unstably, while the delay of the APSM and Scheme-1 increases slowly. That is, the adaptive parameters of the APSM can extend the stability region of the network.

Figure 5

Comparison of Average Delay of Data Frames Under PAR Traffic.

(A) ρ increases when c = 3. (B) c increases when ρ = 0.3. (C) λ_i = 20 ms, ρ increases with c.

Second, we study the delay of data frames of the APSM in a network with many stations under PAR traffic. In Figure 5B, the total traffic is as light as ρ = 0.3. The delay in the APSM is obviously larger than that in the PSM; however, their differences are controlled within 110 ms. At the same time, Scheme-1 only slightly outperforms the PSM in terms of delay. That is, the adaptive backoff timer has a limited influence on delay when the traffic is light. Figure 5C presents the case with multiple stations when the traffic increases with c. When the traffic is light (i.e., c ≤ 10), the APSM has the largest delay, which is <165 ms, while Scheme-1 and the PSM has a similar delay. When the traffic becomes heavy with c, the delay of Scheme-1 increases sharply while the delay of APSM is smooth. That is, the adaptive BI and LI_i in the APSM are helpful to control the delay of data frames when the traffic is not light.

Therefore three schemes are not suitable to traffic-heavy applications. However, the APSM can extend the stability region of the network and support more traffic. The increase of delay due to the APSM is acceptable when the traffic load is not very heavy. Moreover, the traffic under EXP and CBR distributions has the similar tendency.

Finally, we compare the throughput of stations among the three schemes. The APSM has no bad effect on the total throughput compared with the PSM. For example, as shown in Figure 6 with light PAR traffic in multiple symmetrical stations, the throughput of the APSM is slightly higher than that of the PSM and Scheme-1 in most cases. The other network conditions and traffic patterns have the similar tendency.

Figure 6

Comparison of Throughput, c Increases When ρ = 0.3, PAR Traffic.

4.3 Effects of the Power Consumption Model on the APSM

The power profile of a wireless device has a great impact on the performance of the energy-saving scheme using sleeping [22]. This profile includes the power consumption in transmission, reception, idle mode, sleeping mode, and mode transition (in which clients wake up from sleeping), as well as the wake-up time. Additionally, the energy consumed on the client’s wake-up is the product of wake-up power³ and wake-up time. The set of the above parameters are defined as a power consumption model in this article.

In this subsection, we compare the five power consumption models listed in Table 4, which are measured or applied in different references. We adopt model E in the previous simulations, which is comparable to the hardware characteristics of many wireless interface cards. In model E, the ratio of transmission power to reception power is approximately 120%, which is similar to the ratio in ORiNOCO 11a/b/g ComboCard [25] and CISCO AIRONET 350 series WLAN adapter [8]. The reception power is the same as the idle power, just like in CISCO AIRONET 11a/b/g Wireless cardbus adapter [9]. Moreover, its sleep power is about an order of magnitude lower than its idle power as R_I/S = 2779%, which are common in many popular wireless network interface cards. Here, R_I/S is the ratio of idle power to sleeping power.

Next, the effects of these models on the performance of the APSM are studied. We find that the improvements of the APSM over the PSM mainly depend on the idle/sleeping ratio R_I/S and the network utilization ρ. For example, in Table 5, three stations under the EXP traffic λ₁ : λ₂ : λ₃ = 1 : 2 : 3 are discussed. First, the advantage of the APSM decreases when the traffic increases, as the energy efficiency index η_ee decreases with ρ for the same power model. For the same traffic level, η_ee is the least in model A and highest in model E, whose order is the same as the order of R_I/S. That is, the improvements of the APSM over the PSM increase with R_I/S in general.