MULTIPLE-INPUT-MULTIPLE-OUTPUT

Transcription

1 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS 1 Power Management of MIMO Network Interfaces on Mobile Systems Hang Yu, Student Member, IEEE, Lin Zhong, Member, IEEE, and Ashutosh Sabharwal, Senior Member, IEEE Abstract High-speed wireless network interfaces are among the most power-hungry components on mobile systems. This is particularly true for multiple-input-multiple-output (MIMO) network interfaces which use multiple RF chains simultaneously. In this paper, we present a novel power management solution for MIMO network interfaces on mobile systems, called antenna management. The key idea is to adaptively disable a subset of antennas and their RF chains to reduce circuit power consumption, when the capacity improvement of using a large number of antennas is small. Antenna management judiciously determines the number of active antennas to minimize energy per bit while satisfying the data rate requirement. This work provides both theoretical framework and system design of antenna management. We first present an algorithm that efficiently solves the problem of minimizing energy per bit and, then offer its n-compliant system designs. We employ both MATLAB-based simulation and prototype-based experiment to validate the energy efficiency benefit of antenna management. The results show that antenna management can achieve 21% one-end energy per bit reduction to the front end of the MIMO network interface, compared to a static MIMO configuration that keeps all antennas active. Index Terms Antenna management, energy per bit, multipleinput-multiple-output (MIMO) network interface, power management. I. INTRODUCTION MULTIPLE-INPUT-MULTIPLE-OUTPUT (MIMO) technologies are considered as a leading candidate for the next-generation wireless broadband, due to their capability to significantly increase link capacity [1] [3]. They have been adopted by current and emerging mobile wireless standards such as n, WiMAX, and LTE. The key idea of MIMO is to simultaneously use multiple antennas at the transmitter and receiver. By properly leveraging the multipath effect, MIMO can significantly boost channel capacity compared to a traditional single-input-single-output (SISO) link. Note that in this work we use antenna to refer to the passive antenna and the corresponding RF chain that powers it unless otherwise specified. The simultaneous use of multiple antennas by MIMO network interfaces incurs significant circuit power consumption Manuscript received September 08, 2010; revised January 29, 2011; accepted March 18, This work was supported in part by NSF Awards ECCS/IHCS , CNS/NeTS-WN , CNS/CRI , CNS , and CNS and by the TI Leadership University Program. The authors are with the Department of Electrical and Computer Engineering, Rice University, Houston, TX USA ( hang.yu@rice.edu). Digital Object Identifier /TVLSI due to multiple active RF chains. The circuit power increase is particularly problematic for short-range communication scenarios such as based WLAN where circuit power often dominates the total power consumption of the network interface. Existing work on MIMO mainly focus on improving the channel quality such as data rate under the transmit power budget; little published work has considered the dual problem of reducing power consumption especially the circuit power under a data rate constraint. To address the power challenge, we propose a novel power management solution called antenna management. Antenna management dynamically determines the number of active antennas and transmit power for each active antenna, in order to minimize the energy consumption for delivering each data bit, or achieve minimum MIMO energy per bit, while guaranteeing a required data rate. The key rationale behind antenna management is the mobility of mobile systems. As mobile systems move around, they encounter different propagation environments, which can lead to different capacity benefit from using multiple antennas. Since the circuit power cost of using one active antenna is fixed, different environments may lead to different numbers of antennas to achieve the minimum energy per bit. For example, an indoor environment with rich multipath effect can provide a MIMO channel higher capacity improvement than an outdoor environment with a dominant line-of-sight (LOS) path can. As a result, a larger number of antennas is more likely to be optimal for the indoor environment. We provide both algorithm and system designs of antenna management. The antenna management algorithm efficiently solves the MIMO energy per bit minimization problem with: 1) a pre-built mapping to identify the optimal transmit power and 2) antenna selection algorithms to obtain the optimal number of antennas. Our system design of antenna management is n-compliant. We offer both one-ended and two-ended designs where the former is suitable for a MIMO link between a mobile node and an access point while the latter for that between two mobile nodes. We evaluate the system design of antenna management first with a MATLAB-based simulation, and then a first-of-its-kind prototype of a 4 4 MIMO link using an open-access wireless research platform called WARP. We show the effectiveness of antenna management with simulation and experimental results: on average antenna management can save one-end and two-end power consumption to the front end of the MIMO network interface by 21% and 13% compared to a static MIMO link that always uses all antennas. In summary, in this work we make the following contributions /$ IEEE

2 2 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS Fig. 1. MIMO link with a transmitter and a receiver, each with four RF chains and six passive antennas. We analyze the MIMO energy per bit as function of transmit power and antenna configuration, and subsequently formulate an optimization problem towards its minimization. We offer the antenna management algorithm to efficiently solve the MIMO energy per bit minimization problem. We present n-compliant system designs of antenna management and validate their effectiveness through both simulation and prototype-based experiments. Antenna management can be easily extended to other multiantenna technologies, e.g., beamforming [4], albeit our current design and implementation focus on MIMO. The fundamental motivation of antenna management is that in a multi-antenna system higher link performance is often achieved at the cost of using more antennas and therefore higher circuit power consumption. As a result, one can maximize the energy efficiency as long as the performance requirement is met. The rest of this paper is organized as follows. Section II provides background knowledge on MIMO technologies and discusses related work. Sections III and IV present the theoretical framework of antenna management. Section V offers n-compliant system designs of antenna management. Sections VI and VII provide simulation and experimental evaluations, respectively. Section VIII discusses antenna management in a broader context and Section IX concludes this paper. II. BACKGROUND AND RELATED WORK We first provide background on MIMO technologies, in particular the spatial multiplexing MIMO (SM-MIMO) which improves link capacity by sending independent data streams through multiple antennas. In the rest of this paper we use MIMO to refer to SM-MIMO unless otherwise specified. B. MIMO Channel Model As illustrated by Fig. 1, a MIMO channel can be characterized by a complex matrix, where and denote the number of active antennas in the receiver and transmitter respectively. We adopt the MIMO channel model used by IEEE n [7] where and denote the LOS and non-line-ofsight (NLOS) component of the channel respectively, and is the Ricean factor. The elements in are independent random variables of circular symmetric complex normal distribution. The elements in are all one multiplied by a phase shift where is a random variable uniformly distributed in [0, ). The Ricean factor indicates the propagation condition of the channel, i.e., how dominant the LOS component is compared to the NLOS component. By varying, the model can fit channels with various fading distributions. For example, models ideal Rayleigh fading and models ideal Ricean fading. The change of and can yield channel variation. The above model is normalized by the path loss of each sub-channel. Given a constant distance between the transmitter and receiver, the channel fluctuation due to small-scale node movement can be modeled as above. For clarity, we refer to the change of and as small-scale fading, and the change of path loss as large-scale fading. C. MIMO Channel Capacity We consider a frequency-flat MIMO channel to calculate its capacity. The instantaneous channel capacity can be calculated [1] as A. MIMO Link Architecture Fig. 1 shows a representative architecture of a MIMO link with a pair of transmitter and receiver. A realistic MIMO transceiver usually operates in a half-duplex manner thereby can be either the transmitter or the receiver. A MIMO transceiver can allow more passive antennas than RF chains and employ antenna selection techniques [5], [6] to determine the optimal subset of antennas, i.e., which four out of the six in Fig. 1. Each pair of transmit antenna and receive antenna forms a sub-channel between the transmitter and receiver, and these sub-channels collectively constitute the MIMO link. where is the channel bandwidth, the covariance matrix of transmit signal satisfying the transmit power constraint the total transmit power across all transmit antennas, the noise power at the receiver, and an unit matrix. Note that the maximization is performed over all qualified covariance matrices with. The transmitter needs to know the channel matrix to achieve the capacity by optimally allocating the transmit power [8]. The capacity is the upper bound of the error-free data rate allowed by the MIMO channel. Noticeably,

3 YU et al.: POWER MANAGEMENT OF MIMO NETWORK INTERFACES ON MOBILE SYSTEMS 3 it depends on both the channel, i.e., and, and the MIMO transceiver configurations, i.e.,, and. Higher transmit power and more active antennas can increase the channel capacity, regardless of [1]. However, both will meanwhile increase the transceiver power consumption. A key observation that motivated this work is that under many circumstances more antennas will yield little improvement on the channel capacity, e.g., when the channel has a large so that the sub-channels are highly correlated. Under these circumstances it may not be energy-efficient to employ a large number of antennas. D. MIMO Power Model The power consumption of a MIMO link,, includes that of the transmitter and that of the receiver. includes the power consumed by all the power amplifiers and that by all the other transmitter circuitry [9]. Since the power amplifiers are usually identical, only depends on the total transmit power. Moreover, the circuit power consumption can be divided into that contributed by each active transmit RF chain and that by the circuit shared by all active transmit RF chains. Therefore, we approximate as where is the number of active antennas in the transmitter, and is the drain efficiency of the power amplifier. Similarly, we approximate as where and represent the power consumed by each receive RF chain and that by the shared receive circuit, respectively. A MIMO transceiver can monitor the channel when it is idle. Let denote the power consumption of an idle MIMO transceiver. Clearly, is constant during the idle period. E. Related Work Next we discuss work related to antenna management in three directions. 1) MIMO Capacity Maximization: Existing MIMO capacity maximization works improve the capacity of the MIMO link with a given number of antennas and transmit power budget, while antenna management solves its dual problem: finding the optimal number of antennas to minimize energy per bit under a data rate constraint. One popular approach for MIMO capacity improvement is to employ more antennas than RF chains and dynamically select a subset of antennas to maximize capacity, or antenna selection [5], [6], [10], [11]. Its goal can be considered to minimize energy per bit given a fixed number of active antennas. On the other hand, antenna management further allows the number of active antennas to vary according to the channel condition and data rate requirement. From this perspective, antenna selection is an inherent part of antenna management. Another popular approach for MIMO capacity improvement is to optimally allocate transmit power to each active antenna, or spatial power allocation [8], [12], [13]. However, spatial power allocation seeks to improve the link capacity without considering circuit power consumption. By allocating transmit power across all active antennas and meanwhile optimizing the number of active antennas, antenna management naturally includes spatial power allocation. 2) Rate Adaptation: Antenna management can be regarded as a realization of signal-to-noise ratio (SNR)-triggered rate adaptation for MIMO links. Rate adaptation techniques [14] [16] aim to find the data rate that achieves the best throughput as channel condition varies. SNR-triggered rate adaption exploits the received SNR to estimate channel capacity and sets the data rate accordingly. Similarly, antenna management leverages the channel matrix to estimate the MIMO channel capacity and alters data rate by changing the number of active antennas. 3) Energy-Efficient MIMO System Design: Only very recently, the design of energy-efficient MIMO systems has attracted attention. Early results of our work were presented in [17] including an n-compliant prototype and experimental validation. The authors of [18] propose an adaptive MIMO system that can switch between MIMO and SIMO for energy saving. At the mobile end, either one or two transmit antennas are used according to the data rate at the mobile end and the load utilization at the base station. Our work departs from [18] in three ways. First, we have taken into account both ends of communication in two-ended management while the authors of [18] only considered the transmitter energy efficiency. Second, antenna management can be effective for any MIMO configuration, irrespective of the number of RF chains and antennas while the authors of [18] only considered MIMO systems with at most two antennas. Last and most importantly, we have extensively studied system designs and implementation issues of antenna management with both simulation and experimental results while the work in [18] is limited to theoretical analysis and simulation. There have been several experimental studies since [17] and [18]. The authors of [19] employed a different approach to realize antenna management with commodity n chipsets. They use a polling-based method to select the best configuration. Their solution only works for the transmitter, and finds the antenna configuration that can deliver the required data rate with the lowest power consumption, which is not necessarily the most energy-efficient considering idle period. For example, when the MIMO transceiver can enter a power-saving mode during idle time, transmitting as fast as possible and then staying in a power-saving mode can be more energy-efficient. In contrast, antenna management not only addresses both transmission and reception but also automatically considers the use of power-saving mode and idle period by minimizing energy per bit. The authors of [20] experimentally characterized the power consumption of commodity n transceivers with observations similar to ours: the optimal MIMO configuration depends on traffic patterns.

4 4 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS TABLE I CATEGORIZATION OF MIMO ENERGY PER BIT MINIMIZATION III. MIMO ENERGY PER BIT MINIMIZATION In this section we analyze the energy per bit minimization problem for a MIMO link. The question we try to answer is: given the channel matrix, what is the number of active antennas and transmit power that yield minimum energy per bit? A. Objective Function The objective function, the MIMO energy per bit, can be calculated as the power consumption divided by the data rate. We have two observations regarding and. First, the two ends of a MIMO link can be either a pair of battery-powered mobile nodes, or an infrastructure node and a mobile node. Because energy efficiency is only important to mobile systems, can be either the power consumption of both ends or that of a single end. Second, the data rate supported by the MIMO channel is function of the number of antennas at both ends. In practice, one or both ends may allow antenna management. These two observations lead to nine cases of the MIMO energy per bit minimization problem, depending which end allows antenna management and which end desires energy efficiency optimization. When both ends are mobile nodes, we have ; when only one end is mobile, we have and for the mobile node as a transmitter and a receiver, respectively. In addition, when both ends allow antenna management, we have ; if only one end supports antenna management, we have for the transmitter and for the receiver. Therefore, there can be nine cases, summarized in Table I: the rows refer to at which end the energy efficiency should be optimized and the columns to which end supports antenna management. A further examination reveals that only Case 1, Case 5, and Case 9 are nontrivial (See Appendix A). For clarity, we name Case 1 as two-end energy per bit minimization, and Case 5 and Case 9 as one-end energy per bit minimization. Their objective functions can be written as follows. Case 1: Case 5: Case 9: B. Constraint We add an important constraint to the optimization problem: being the minimum required data rate, because the optimization variables and have a direct impact on the data rate and wireless links usually have a data rate requirement. C. Optimization Variables We next look into the optimization variables. Without loss of generality, we use Case 1 for discussion. First, we observe that given and, there exists a finite optimal transmit power,, that yields minimum. This observation ensures the need of optimization for and we provide the detailed proof in Appendix B. Second, we combine other optimization variables as one, namely the antenna configuration,, which includes not only the number of antennas, i.e., and, but also which subset of antennas is active. Apparently, each yields an unique channel matrix and thereby a unique optimal transmit power. IV. ALGORITHMIC DESIGN OF ANTENNA MANAGEMENT We next provide an efficient solution to the optimization problem formulated in Section III. The problem is non-trivial to solve for the following two reasons regarding and, respectively. Given and, no closed-form formulation of the optimal transmit power,, is obtainable. The number of possible antenna configurations increases exponentially with the number of antennas so that a bruteforce method to identify yields the complexity of. Antenna management leverages two key techniques to tackle the above challenges. First, it identifies with mappings built offline. For each pair of and, antenna management employs multiple mappings to cope with large-scale channel fading introduced by significant movement of the mobile node. Second, for a small number of antennas, antenna management enumerates all the antenna configurations to find, i.e., the optimal and the optimal subset of antennas; for a large number of antennas, it leverages existing antenna selection algorithms. The overall algorithm is summarized in Algorithm 1 and we next describe the two techniques. 1) Pre-Built Mapping: We observe that, without considering is primarily determined by the dimension of,

5 YU et al.: POWER MANAGEMENT OF MIMO NETWORK INTERFACES ON MOBILE SYSTEMS 5 Fig. 2. Antenna management is implemented inside the n WNIC. It obtains data rate constraint from MAC and channel estimation from PHY, and then selects the optimal antenna configuration. i.e., and, under small-scale channel fading. In other words, for fixed and, one shall expect approximately the same, when the channel condition does not change much. Therefore, a mapping from each to can be built offline using either synthetic or measured channels. In our work, we generate synthetic channels with the channel model described in Section II and obtain measured channels using the same setting as that used in Section VII. We first use a brute force method to compute for each synthetic/measured channel matrix with various and.for each pair of and, we use the median of the computed for all channels as the approximation to the optimal transmit power,. To see how good is, we compare the MIMO energy per bit using and for a large number of synthetic and measured channels (1000 each). The results show that the MIMO energy per bit using is only 1.3% and 1.6% higher than that using, for synthetic and measured channels, respectively. This overhead is negligible compared to the energy per bit reduction by antenna management. Similarly, one can build a mapping from to offline because the mapping is largely independent from under small-scale channel fading (see Appendix B). 2) Antenna Selection: In addition, we convert the procedure of finding into multiple steps where each step can be solved by existing antenna selection algorithms. Given and, minimizing is equivalent to maximizing since the power consumption is constant. Therefore, it turns into a capacity-maximization-based antenna selection problem where existing efficient algorithms, e.g., [11], can be straightforwardly leveraged. As a result, we only need to identify the optimal number of antennas, with a complexity of. V. SYSTEM DESIGN OF ANTENNA MANAGEMENT We next provide the system design of antenna management. We choose the MIMO-based WLAN standard, IEEE n, as the operating protocol due to its commercial availability. A. MIMO-Based n n supports MIMO with up to four RF chains integrated in the MIMO network interface. More than one passive antenna can be attached to each RF chain to enable antenna selection. Each RF chain together with its selected passive antenna is responsible for sending a spatial stream. A single frame from MAC can be broken up and multiplexed across multiple spatial streams and then reassembled at the receiver. The number of spatial streams, or active antennas, is allowed to be dynamic. The PLCP header specifies this number so that the receiver can correctly decode the original signals. B. Design Overview We present two n-compliant designs of antenna management. The first design is one-ended targeting the mobile node in a legacy network. The mobile node only considers its own energy efficiency. The second design is two-ended, with antenna management at both ends to minimize energy per bit for the MIMO link. It is desirable when both ends are energyconstrained, e.g., two mobile nodes in an ad-hoc network. The one-end and two-end energy per bit minimization in Section III are the theoretical foundations for the above two designs. Both designs intend antenna management to be implemented inside the MIMO WNIC, as illustrated by Fig. 2. Antenna management obtains the channel estimation from PHY and data rate constraint from MAC, and then choose the optimal MIMO configuration. We next discuss the issues common to both oneended and two-ended antenna management, and then address those specific to each of them. 1) Setting the Data Rate Constraint: While n supports very high data rates, wireless interfaces on mobile systems usually experience much lower rates, primarily due to the bottleneck in the rest of the network or system, e.g., DSL link or applications like VoIP with low rate of packet supply [21]. Therefore, the minimum constraint of data rate,, should be set by upper layers that can assess the application requirements or performance bottleneck. The required data rate from upper layers usually varies much slower than the rate variation supported by n, i.e., on a per-frame basis for the latter. Therefore, the data rate of each frame should be constrained in a way that the average data rate over a short period meets the constraint. As a result, we set the per-frame data rate constraint,, according to

6 6 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS Fig. 3. Mapping from the number of antennas to the optimal transmit power. Fig. 4. Mapping from the effective data rate to the optimal transmit power. where is the index of the frames to which antenna management is applied, the experienced average rate of frame 1to, and the weight given to current frame. Naturally, if each frame has approximately the same length, one should set to. 2) Estimating the MIMO Channel: It is crucial that the channel matrix is known by the antenna managed mobile nodes in order to estimate the channel capacity. Channel estimation by the receiver is known to be straightforward; there are two ways to estimate the channel by the transmitter, namely closed-loop channel estimation and open-loop channel estimation. Closed-loop channel estimation requires the transmitter to send training symbols to the receiver and the receiver calculates and then feeds it back to the transmitter. This method maintains accuracy but incurs overhead because of the feedback of. Open-loop channel estimation, instead, leverages channel reciprocity, lets the transceiver estimate in receive mode and assumes it as the channel for next frame transmission. It is effective only when the communication is by-directional n supports both closed-loop and open-loop channel estimation. The PLCP preamble of each frame contains training symbols that are leveraged by antenna management to achieve channel estimation as we will show later. 3) Identifying the Optimal Antenna Configuration: In Section IV we have presented an efficient algorithm to identify the optimal transmit power and the optimal antenna configuration. That is, we first establish a mapping from the number of antennas to, as well as a mapping from the data rate to. To cope with large-scale channel fading, the mappings should be updated when the path loss has considerable variation. One can simply check the RSSI available in n to detect the path loss variation, since the RSSI change due to small-scale fading is relatively small. Fig. 3 shows an example of the mapping from to and Fig. 4 shows that from to. Also note that the transmit power in realistic n transceivers is usually not continuously available so that the mapping can be further simplified. To get, we need to estimate the data rate for each. While one can calculate the channel capacity using as the upper bound of data rate so that minimize the lower bound of energy per bit, the solution can be sub-optimal under non-continuously available rates. Therefore, specifically for n, we also need a mapping from the channel matrix to the achievable data rate. One can calculate the channel capacity and accordingly find the most appropriate data rate. Such procedure is already used in many rate adaptation works, e.g., [16], [22]. We also note that additional channel properties such as the bit error rate (BER) can be taken into account during this procedure and they are complementary to antenna management. C. One-Ended Antenna Management Next we present the unique design principles of one-ended antenna management and target two-ended management in the next section. One-ended antenna management leverages open-loop channel estimation. It is triggered when the MIMO transceiver receives a frame in the receive mode. The default configuration of the receive mode has all antennas active, which is essential for the transceiver to use open-loop channel estimation. The transceiver then leverages the estimated channel matrix to obtain the achievable data rate, gets the optimal transmit power for each antenna configuration, and then identifies the one with minimum energy per bit. The open-loop channel estimation is effective because of the acknowledgement mechanism intrinsic to : a receiver immediately sends back an ACK frame to the transmitter to acknowledge the reception of a data frame, which provides the transmitter a free opportunity to estimate the channel. Moreover, provides a carrier sensing mechanism with the RTS/CTS frame exchange before data frame transmission. While RTS/CTS is only used for large data frames, it guarantees the effectiveness of open-loop channel estimation, especially for non-continuous transmission where a long interval may exist between the last ACK frame and the current data frame. D. Two-Ended Antenna Management Two-ended antenna management leverages closed-loop channel estimation, but does not incur additional frame transmission due to two key reasons. First, since each data frame actually grants the receiver a free opportunity to estimate using the training symbols in the PLCP preamble (see Section V-A), no extra training frames from the transmitter are needed. While all transmit and receive antennas need to be active to deliver the training symbols, the resultant energy overhead is negligible since the training symbols are much shorter compared to the frame body. Second, the receiver only needs

7 YU et al.: POWER MANAGEMENT OF MIMO NETWORK INTERFACES ON MOBILE SYSTEMS 7 to send back the optimal antenna configuration instead of to the transmitter. Once the transmitter knows, it can easily find the optimal transmit power using the pre-built mapping. Since can be encoded as indexes, e.g., a4 4 MIMO link has 256 possible configurations and can be uniquely represented by 8 bits, it is allowed to be included in the ACK frame sent immediately after the data frame. We note that to let the receiver identify, the power profile of the transmitter, including and mapping for, needs to be known by the receiver. However, these parameters can be exchanged by the two ends in advance of data transmission, and such exchange is needed only once. E. Overhead It is important to note that both one-ended and two-ended antenna management can be almost overhead-free in terms of latency and energy. First, modern transceivers incur very small latency for switching on/off a entire RF circuitry, e.g., 50 s for the MAX2829 transceiver [23] used in our prototype. Second, our measurement showed that the energy overhead can be readily compensated by powering off the transceiver for as short as 100 ns, while frame transmission may last several milliseconds. Last and most importantly, antenna management does not incur additional mode switching besides the regular transition between transmit and receive modes, which is inevitable even without antenna management since n transceivers are half-duplex. VI. SIMULATION-BASED EVALUATION In this section, we use a MATLAB simulation to evaluate antenna management under synthetic channels. A. Simulation Setup We employ MATLAB to simulate a MIMO link that includes two identical n-like transceivers, denoted as Node 1 and Node 2. Each transceiver has four RF chains and antennas, in accordance to n. To reveal the maximal benefit of antenna management, we go beyond current n transceivers and assume that both the transmit power and the data rate can be arbitrarily chosen. The n-compliant transmit power and data rate settings will be assumed for our prototype-based evaluation in Section VII. Parameters in the power model are set as follows: 48.2 mw, 62.5 mw, 50 mw. Those numbers are chosen according to the model in [9] and very close to recently reported realizations of n transceivers such as [24] (without baseband power). We use two traffic patterns, continuous and intermittent, to represent different frame arrival rates. For continuous traffic we assume that frames from upper layers arrive at an extremely high rate so that the MIMO transceivers are always engaged in transmitting or receiving, i.e., idle period never appears. This traffic represents applications such as FTP which supply data at a much higher rate than the maximum rate supported by the transceiver. Users can specify the minimum data rate that must be satisfied by antenna management. Intermittent traffic, in contrast, has a much lower frame supply rate and may introduce idle period between successive frames, during which the transceiver enters idle mode. Intermittent traffic represents applications such as VoIP which generate sparse traffic and a relatively lower data rate. The data rate constraint is determined by the application. While the distribution of inter-frame interval can be arbitrary, e.g., exponential, we assume a constant interval in our simulation for simplicity. Therefore, the intermittent traffic is completely periodic. We note that some applications such as HTTP usually have both continuous and intermittent traffic patterns: data frames from applications usually arrive in bursts; while each burst can be considered as a continuous traffic of multiple frames, the bursts come intermittently. It is also worth noting that the inclusion of idle period energy does not change the structure of the MIMO energy per bit minimization problem so that antenna management is still valid under intermittent traffic. We provide elaborated explanation regarding this in Appendix C. For both continuous and intermittent traffic, we explore five scenarios for evaluation, using synthetic channels based on the channel model in Section II. For the first four scenarios we use a constant Ricean factor that equals 0, 1, 10, 100, respectively, to represent different fading distributions; for the last scenario we assume is random. We do not consider large-scale fading in the simulation. All the channels are reciprocal with a 20 MHz bandwidth. We also assume the channel coherence time is longer than the frame duration so that the per-frame adaption of antenna management can perfectly cope with channel variation. In fact, real-world channels usually exhibit longer coherence time according to our channel traces in Section VII. We use a static configuration with all antennas active as the baseline to evaluate the energy efficiency of antenna management. For two-ended management we measure the average energy per bit of the MIMO link for 1000 data frames, with the first 500 frames transmitted by Node 1 and the second 500 frames by Node 2. For one-ended management, we measure the energy per bit of Node 1 which implements antenna management and transmits 1000 data frames to Node 2. For both measurements we employ four different data rate constraints: 0, 100, 200, and 300 Mbps. B. Simulation Results We first show the simulation results under continuous traffic and then those under intermittent traffic. 1) Continuous Traffic: Fig. 5(a) and (b) shows the MIMO energy per bit reduction by two-ended and one-ended antenna management, respectively. We make several key observations from Fig. 4(a) and (b). First, the energy per bit achieved by antenna management is strictly no larger than that of the static configuration. Twoended and one-ended antenna management achieves 13% and 21% energy per bit reduction, respectively, where the scenario with a random is considered for antenna management to offer average performance. Second, when the data rate constraint increases, the energy per bit reduction by antenna management drops. This is because higher data rate requires more active antennas so that the optimal configuration given by antenna management gradually approaches the static one as data rate constraint increases.

8 8 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS TABLE II ROBUSTNESS OF ANTENNA MANAGEMENT UNDER INACCURATE CHANNEL ESTIMATION of idle energy consumption. When the data rate constraint is low, i.e., the traffic is sparse, both transceivers spend very little time in transmitting or receiving. Therefore, the transmitting and receiving optimization by antenna management improves energy efficiency trivially. For very high data rate constraints, intermittent traffic indicates similar results as continuous traffic does. We must note that many have addressed the energy efficiency issue of idle periods, e.g., the authors of [21]. Such work is complementary to antenna management. Fig. 5. MIMO energy per bit reduction by antenna management in the simulation-based evaluation. (a) Two-ended, continuous traffic. (b) One-ended, continuous traffic. (c) Two-ended, intermittent traffic. (d) One-ended, intermittent traffic. Third, under a relatively low data rate constraint, e.g., less than 200 Mbps, antenna management becomes more effective with a larger. This is because under channels with a relatively large, adding more antennas brings marginal capacity improvement but incurs fixed additional power consumption. Therefore, the static configuration using all antennas can be very inefficient. On the other hand, under a high data rate constraint, antenna management reduces energy per bit more for channels with a smaller. This is because using the same number of antennas, a channel with a smaller is able to yield higher channel capacity so that more likely to meet the data rate constraint. 2) Intermittent Traffic: We then evaluate antenna management under intermittent traffic, shown in Fig. 5(c) and (d). While most of our observations for the continuous traffic are also applicable to the intermittent traffic, the latter has unique characteristics. First, the peak energy per bit reduction for intermittent traffic is higher than that for continuous traffic. For example, with a random, one-ended antenna management reduces energy per bit by 37% for intermittent traffic versus 21% for continuous traffic. This is because antenna management under intermittent traffic also reduces energy consumption in idle periods. Recall that for intermittent traffic we assume a fixed interval between successive frames. Therefore, using fewer antennas to extend active transmitting or receiving duration can result in shorter idle period thereby less idle energy consumption. Second, under intermittent traffic there exists certain data rate constraint under which antenna management is most effective, e.g., 200 Mbps for two-ended management and 100 Mbps for one-ended management. Again, this is due to the consideration C. Robustness Against Channel Estimation Error Next we use simulation results to demonstrate that antenna management offers great energy saving even when the channel estimation is reasonably inaccurate. We assume one-ended management for simplicity while the methodology and answer also hold valid for two-ended management. The effectiveness of antenna management depends on the channel estimation accuracy. One-ended management employs open-loop channel estimation so that the estimated may greatly deviate from its actual value when the received SNR is low. We examine the impact of channel estimation error on antenna management by assuming different received SNR in the simulation and comparing the theoretically optimal configuration with the one identified by antenna management. Table II shows the probability that antenna management offers the theoretically optimal configuration and the average energy per bit increase for four different SNR levels, from 0 to 20 db. Even when the SNR is very low, i.e., 0 db, antenna management consumes only about 1% more energy per bit than the theoretically optimal configuration does, which is negligible in comparison to the reduction achieved from a static configuration. VII. PROTOTYPE-BASED EVALUATION We next measure the performance of antenna management with a prototype implementation under realistic channels. Our results have confirmed the energy efficiency benefit of antenna management and meanwhile demonstrated its feasibility in practical MIMO network interfaces. A. Prototype Implementation We have implemented a prototype of antenna management using an open-access wireless research platform, WARP [25]. We build a 4 4 MIMO link consisting of two WARP nodes each with four RF daughterboards. Each daughterboard has a MAX2829 transceiver [23] connected with one passive antenna. We use a Lenovo ThinkPad T400 laptop to interact with the two

9 YU et al.: POWER MANAGEMENT OF MIMO NETWORK INTERFACES ON MOBILE SYSTEMS 9 WARP nodes via WARPLab, a MATLAB-based interface that enables the laptop to send commands to and collect measurements from WARP nodes. Due to space limitation, in the following we only report the measurement results for one-ended management, which is sufficiently informative to show the feasibility and practical benefit of antenna management. We have emulated an n WNIC in WARPLab, with realization of the PHY functionality related to antenna management, including PLCP preamble/header, modulation and coding, and transmit power control. In accordance to n, we implement different modulation schemes including BPSK and MQAM, as well as different coding rates including 1/2, 2/3, 3/4, and 5/6. Therefore, the data rate is not continuously available. In addition, we set the resolution for adjusting the transmit power as 1 dbm, where the transmit power is no greater than 20 dbm (100 mw) consistent with n. Since our prototype-based evaluation is limited to a single MIMO link, we omit most MAC functionality and only emulate MAC frame generation and collection. We also highlight that our prototype uses a simple searching method to find the optimal antenna configuration. This is because one-ended antenna management with four RF chains and antennas used in our prototype only yields 16 possible configurations so that the antenna selection algorithm is unnecessary. B. Experimental Setup Fig. 6 shows our experimental setup. The setup involves two WARP nodes: one as the mobile node implementing antenna management and the other as the access point with all antennas active. The radio is at 2.4 GHz, and the channel bandwidth is 20 MHz. We choose two typical indoor environments: one has a dominant LOS path between two nodes (called the LOS scenario) and the other one does not have any LOS paths (called the NLOS scenario). For the LOS scenario we put the two WARP nodes in the same office, and for the NLOS scenario we put two WARP nodes in two adjacent offices apart from each other by a wall. While the WARP nodes are connected to the Ethernet switch via a wired cable thereby cannot have large-scale movement, we introduce small-scale mobility to the nodes by manually moving them as well as the scatters. Therefore, the channel has small-scale fading. We generate null n data frames at the mobile node and send them to the access point. A null frame contains random binary bits in the frame body, which is used to mimic realistic n frames with application data. In addition, since we have addressed the difference between continuous and intermittent traffic in the simulation, we only employ continuous traffic for the prototype-based evaluation. C. Measurement Results We report measurement results for two different MIMO links: 2 2 and 4 4. For each of the two, we compare antenna management with a static configuration with all antennas active, as adopted in the simulation. To guarantee fair comparison, we evaluate antenna management in real time, while emulate the static configuration offline using the collected channel traces. Since the energy per bit of the static configuration is uniquely Fig. 6. Experimental setup for the prototype-based evaluation of antenna management: one WARP node with antenna management emulates the mobile node, the other WARP node with legacy n emulates the access point, and one laptop controls both nodes as well as collects data. Fig. 7. Energy per bit reduction by antenna management in the prototype-based evaluation. (a) MIMO and (b) MIMO. determined by the channel, one can expect the emulation to offer the same results as real-time evaluation does. The results are presented in Fig. 7. We have shown the energy per bit reduction by antenna management with different data rate constraints for the two MIMO links each in both LOS and NLOS scenarios. While the results for the 4 4 MIMO link echo the simulation results in Section VI, one can see that antenna management becomes less effective for the 2 2 MIMO link. This is due to the fact that a 2 2 MIMO link yields fewer choices of antenna configurations so that the static configuration is more likely to be optimal. It is then apparent that as the number of RF chains and antennas in mobile systems becomes larger and larger with the progress of integrated circuits technology (Moore s Law), higher diversity of the antenna configuration can make antenna management increasingly energy efficient. VIII. DISCUSSIONS We next discuss the limitations and extensions of antenna management. A. Limitations Our work is limited in the following aspects that can be addressed by future works. First, we only considered the energy impact of the MIMO transceiver front-end without that of the baseband. In fact, the number of active antennas also affects the processing load in the baseband, especially for the receiver.

10 10 IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS Typically, the more antennas, the more processing and higher power consumption in baseband. Therefore, antenna management will conserve energy in baseband too by using a smaller number of active antennas. Second, our antenna management algorithm and implementation targets a single link in the network without taking into account the energy efficiency or data rate of other links. In fact, antenna management may negatively affect other links, when these links are time-sharing the channel. For instance, in a typical network with one access point and multiple associated mobile nodes, the mobile nodes with antenna management may occupy the channel for longer time since only a subset of antennas are active with a lower data rate. Consequently, the throughput of other nodes may decrease and their power consumption will increase since they need stay idle for longer time to wait for the shared channel to become free. As a result, antenna management needs further optimization in heavily loaded networks. B. Extension to Cellular Networks Antenna management can be leveraged in cellular systems, especially for the fourth-generation (4G) mobile broadband [26], albeit our design reported in this work targets WLAN. Current base stations in cellular networks have already embraced multi-antenna techniques, e.g., switched-beam systems [27] or beamforming systems [28]. Besides the base station, even the mobile client can employ multiple antennas, in order to increase uplink data rate. One-ended antenna management can be suitable for the mobile client in cellular networks. Nonetheless, one needs to consider important dissimilarities between WLAN and cellular networks. For example, uplink and downlink channels in cellular networks are usually not reciprocal, especially for networks operating in frequency-division-duplex (FDD) mode. Therefore, open-loop channel estimation is no longer effective and one has to use closed-loop estimation. We also note that since cellular channels are more likely to be outdoors where the propagation path is closer to LOS. Recall our results showed that antenna management reduces more energy per bit when the path is closer to LOS. Therefore, one can expect even higher energy efficiency benefit of antenna management for the use in cellular networks. IX. CONCLUSION In this paper, we presented a power-saving mechanism, antenna management, to maximize the energy efficiency of the MIMO network interface on mobile systems. Antenna management adaptively optimizes the transmit power and antenna configuration in order to achieve the minimum energy per bit under a given data rate constraint. We showed that antenna management can be realized with little change to the n protocol to maximize the energy efficiency of a single end or both ends of a MIMO link. Our evaluation using both MATLAB-based simulation and prototypebased experiment demonstrated that antenna management on average can achieve 13% two-end energy per bit reduction and 21% one-end energy per bit reduction. APPENDIX A We discuss the nine cases of the MIMO energy per bit minimization problem in Table I and show that we can reduce them to three non-trivial cases. First, when the power consumption of both ends are considered, optimizing a single end becomes a special case of optimizing both, e.g., one can regard the transmitter optimization as the optimization for both ends with a fixed receiver antenna configuration, and similarly for the receiver optimization. Therefore, Case 2 and Case 3 are inherently included in Case 1. Analogically Case 4 can be broken down into Case 5 and Case 6, and Case 7 can be broken down into Case 8 and Case 9. Second, when the power consumption of a single end is considered, optimization at the other end has no impact on the power consumption and only changes the data rate.for Case 6 and Case 8, the minimization of is equivalent to the maximization of, and this is in fact the capacity maximization problem without considering power consumption. It has been solved by antenna selection techniques and is out of the scope of this work. In other words, Case 4 is equivalent to Case 5, and Case 7 is equivalent to Case 9. To summarize, Case 1, Case 5 and Case 9, are the only three non-trivial cases that do not overlap with each and cannot be solved by existing techniques. APPENDIX B Here we prove the existence of unique that minimizes. Toward this, we only need to prove that there is a single minimum of the objective function. While calculating the second-order derivative of is not straightforward, we can alternatively show the following two facts which equivalently prove our statement: 1) the upper bound of is infinite; 2) has a unique solution in. 1) is obvious given that. 2) can be proven as below. First, we rewrite as where is the rank of and denotes the positive eigenvalues of. Here we use channel capacity to represent data rate for tractability. Then implies (1) where. Now we show that (1) always has a unique solution where. To prove that has a unique solution of in, we can equivalently prove: (i) ; (ii) ; and (iii) in, since is a continuous function. They are respectively proved as follows. (i) because and are the positive eigenvalues of.