Abstract
Rate adaptation, which dynamically chooses transmission rate provided at the physical layer according to the current channel conditions, is a fundamental resource management issue in IEEE 802.11 networks with the goal of maximizing the network throughput. Traditional rate adaptation algorithms for IEEE 802.11n networks do not consider the interference problem, which becomes much more serious due to the rapid deployment of IEEE 802.11n devices and large number of mobile terminals. In this paper, an interference-aware rate and channel adaptation scheme RaCA for intensive IEEE 802.11n networks was proposed. Firstly, RaCA leverages RSSI and CSI information together to measure the current channel conditions at the receiver side. RSSI is a coarse-grained indicator and CSI is a fine-grained indicator. Secondly, a two-stage rate adaptation scheme TSRA was designed, which can quickly adapt to optimal bit rate based on RSSI and CSI information. Finally, a quorum-based channel adaptation algorithm QCA was proposed, which does not need control channel. If channel suffers severe interferences, RaCA calls QCA to choose another channel to work on. Simulation and testbed implementation results demonstrate that RaCA achieves significant throughput gain over SampleLite and Minstrel-HT.
1. Introduction
IEEE 802.11n is the most widely deployed wireless local area network (WLAN). It combines the new physical and link layer enhancements, including channel bonding, MIMO (Multiple-Input Multiple-Output), and frame aggregation [1]. Therefore, IEEE 802.11n can provide a series of bit rates ranging from 6.5 Mbps to 600 Mbps at physical layer. However, how to use these bit rates is not specified in the IEEE 802.11n standard. Rate adaptation that dynamically adjusts the transmission rate, namely, modulation and coding schemes (MCs), according to current channel states, has a great impact on throughput performance for IEEE 802.11n networks than on legacy IEEE 802.11a/b/g networks [2]. In the last decade, researchers have proposed lots of rate selection schemes for IEEE 802.11 a/b/g networks. However, as shown in [3], the rate adaptation algorithms used in IEEE 802.11a/b/g networks turn out to be inefficient when applied in the IEEE 802.11n networks due to different physical and MAC layer technologies. These observations led to new searches on rate adaptation schemes for IEEE 802.11n networks in the last few years. The latest search results implemented a novel hybrid link adaptation scheme called SampleLite [2], which adapts all IEEE 802.11n features including MIMO mode, channel bonding, modulation and coding scheme, and frame aggregation level with varying channel conditions. ARAMIS [3] simultaneously adapts both rate and channel width which obtains a 10-fold increase in throughput over leading IEEE 802.11n rate adaptation contenders. ORS [4] presents a family of rate adaptation algorithms by formulating the rate adaptation problem as an online stochastic optimization problem. The authors in [5] develop a model-driven energy-aware rate adaptation approach which optimizes the energy consumption and can effectively trade off between energy and throughput. EERA [6, 7] implements an energy-based rate adaptation algorithm that trades off between goodput and energy savings at NIC (Network Interface Card). The MiRA [8] uses a novel zigzag rate adaptation scheme, which opportunistically zigzags between intramode and intermode rate options.
All the above rate adaptation algorithms do not consider the interferences problem. However, with the emergence of the IEEE 802.11n standard, more and more APs (Access Points) are deployed in IEEE 802.11 WLANs, and a large number of mobile terminals need to be connected to the Internet through IEEE 802.11 WLANs. Figure 1 shows the number of terminals connected to one AP measured by smartphones in one laboratory building in our campus. We found that more than 100 terminals are connected to one AP at the same time and sometimes can reach 170. Meanwhile, with the massive development of Internet of Things (IoT) applications, a wide variety of embedded IoT devices and sensor nodes need Internet access through IEEE 802.11n networks to communicate, which will produce a large number of data in IEEE 802.11n networks [9, 10]. This makes interference problem much more serious [9, 10] in intensive IEEE 802.11n networks. Therefore, it is vital to take the interferences into account when designing new rate adaptation algorithms in such intensive IEEE 802.11n networks. Particularly, in highly intensive IEEE 802.11n networks [11], it is inefficient to adapt the modulation and coding schemes only [12], and reducing bit rate when heavy collisions happen will seriously reduce the network performance on the contrary [13].
In this paper, a joint rate and channel adaptation scheme called RaCA is proposed, which dynamically chooses appropriate rate and channel based on the current link conditions. Specifically, when the channel-errors are the main factor that affects the communications, the RaCA will adapt bit rate accordingly. Otherwise, the RaCA will choose another different channel to work on when heavy collisions happen.
Our major contributions are summarized as follows.(i)Firstly, a joint rate and channel adaptation scheme called RaCA is proposed using RSSI (Received Signal Strength Indicator) and CSI (Channel Station Information) information for intensive IEEE 802.11n networks.(ii)A two-stage rate adaptation approach named TSRA is proposed. TSRA uses RSSI to choose number of spatial streams, channel width, and guard interval and uses CSI to decide other MCs parameters. TSRA significantly reduces the rate search space from 300 to 10 below.(iii)Without requiring control channel, a channel adaption scheme called QCA is designed using quorum system, which enables efficient rendezvous on a certain channel between transmitter and receiver pair. When communications suffer severe interferences, RaCA adaptively chooses another channel to work on using QCA.(iv)Finally, the performance of RaCA is evaluated through simulations and testbed implementations. Experimental results demonstrate that RaCA outperforms SampleLite and Minstrel-HT in terms of throughput under various network scenarios.
The rest of the paper is organized as follows. The related work on rate adaptation problem in IEEE 802.11n networks is discussed in Section 2. In Section 3, the design details of RaCA are presented. The detailed performance evaluation and simulation and analysis for RaCA are described in Section 4. Finally, Section 5 concludes the paper.
2. Related Work
Rate adaptation is one of the hottest research subjects in recent years for IEEE 802.11 networks, and a number of rate adaptation protocols are proposed in the literature. Existing research results on rate adaptation can be classified into two categories: IEEE 802.11a/b/g-based and IEEE 802.11n-based. IEEE 802.11a/b/g-based rate adaptation schemes have been well surveyed in [14, 15]. Here, we only review the recent research results on IEEE 802.11n-based rate adaptation protocols, which are relevant to RaCA.
Pefkianakis et al. [8] investigated the rate adaptation problem in IEEE 802.11n networks and discovered the nonmonotonicity between frame error rate and bit rate across different MCs. Based on this observation, they proposed a rate adaptation scheme called MiRA that zigzags across different MIMO modes to search for the rate that provides the maximum goodput. Minstrel-HT [16] is a rate adaptation mechanism for the IEEE 802.11n/ac standard based on Minstrel [17]. Minstrel-HT is based on the approach of probing the channel randomly to dynamically learn about working rates that can be supported. Minstrel-HT chooses bit rate to provide the highest throughput for data transmission. Minstrel-HT reduces the number of streams to use lower bit rate if the selected bit rate proves to be too lossy. Minstrel-HT is the default new rate selection scheme implemented in ath9k [18], which works well. ESNR [19] used the concept of effective SNR (Signal-to-Noise Ratio) based on CSI to develop an OFDM (Orthogonal Frequency Division Multiplexing) receiver model and proposed a rate selection algorithm. ESNR computes the highest bit rate configuration based on CSI information, which is expected to successfully transmit data packets. Kriara et al. [2] used RSSI measured at sender-side to design SampleLite, a new hybrid link adaptation algorithm. SampleLite was based on the following observation: each MC exhibits monotonicity with RSSI that is measured at sender-side.
Differently from the schemes outlined above, the work in [5, 6] focused on energy-aware rate adaptation algorithms. In [6], the authors proposed a new rate adaptation algorithm termed EERA that trades off goodput for energy savings at an IEEE 802.11n client NIC. EERA searches for the MIMO setting consuming less per-bit energy, rather than achieving higher goodput. The authors in [5] observed that for a fixed number of antennas, the energy consumed by transmitting or receiving a packet is proportional to the expected transmission time (ETT), and the slope of the energy consumption versus ETT lies on the number of antennas being used. Based on these observations, they designed a simple yet precise model to predict the energy consumption when a specific rate is used and developed a model-driven rate adaptation scheme based on the model to choose the rate that optimizes energy consumption.
Recently, researchers have proposed different rate adaptation schemes for different application requirements [20–23]. MuDRA [20, 21] proposed a multicast dynamic rate adaptation algorithm, which achieved tradeoff between stability for multimedia applications and rapid response to channel conditions. Some nodes in MuDRA collect information using a lightweight protocol and the rate selection response time is adjusted dynamically. LLRA [22] implemented a latency-aware rate adaptation scheme for delay-sensitive applications. TiM [23] designed a novel three-dimensional modulation scheme, timeline modulation, which includes time, amplitude, and phase domains. InFRA [24] designed an interference-aware physical/FEC (Forward Erasure Correction) rate selection framework for multicast video over WLANs. InFRA considered the interference for the first time in rate adaptation and FEC problems. STRALE [25] implemented a joint bit rate and frame aggregation length adaptation in mobile WLANs scenarios.
All the existing rate adaptation schemes except InFRA do not consider the interferences. Particularly, these rate adaptation schemes do not discern the causes of packet transmission failures, interferences or channel-errors. Without differentiating the causes of transmission failures and reducing the transmission rate blindly, this will degrade the network performance on the contrary. InFRA [24] uses RSSI and cyclic redundancy check (CRC) error notifications to infer the cause of the packet losses. In this paper, we propose a joint rate and channel adaptation scheme RaCA based on the RSSI and CSI information, which only lowers the transmission rate due to channel-errors, and will select another channel to work on if the link suffers severe interferences. RaCA adopts BLMon [26] to accurately differentiate the causes of transmission failures using frame aggregation and block acknowledgements information.
Differences from Conference Version. Parts of our work have been published in the International Conference on Advances in Information Technology (IAIT) 2016, which is entitled as RaCA: A Joint Rate and Channel Adaptation Scheme for Dense 802.11n Networks [27]. As compared to the conference paper, a lot of technical details are revised and enhanced. The main improvements are as follows. Firstly, the latest research results on rate adaptation problem for IEEE 802.11n networks are introduced, which further demonstrates the motivation for designing RaCA. Secondly, the design details of RaCA have been substantially improved, which lies in the following aspects. Figure 2 is added in Section 3.1, which describes clearly the modules and interactions in RaCA. We have described the design idea about TSRA scheme in detail in Section 3.2. The testbed experiment analysis on RSSI and CSI is added in Section 3.2, which further demonstrates the motivation for designing TSRA, namely, why we use RSSI to choose number of spatial streams, guard interval, channel width, and CSI to determine other parameters. The RSSI estimator is also added in Section 3.2, which enhances the foundation of LSD design. The TSRA algorithm and the definition of variables used in TSRA algorithm are added in Section 3.3. Also, we have explained how to choose the thresholds for variables in TSRA algorithm in our simulations in Section 3.3. These improvements have substantially enhanced the design details of TSRA. Section 3.4.1 is added in Section 3.4, which reviews the background information on quorum system used in QCA. Also, the QCA design steps and the pseudocode for QCA algorithm are added in Section 3.4, and a channel adaptation schedules example is added in Figure 8. These improvements have greatly enhanced the design details of QCA. Finally, the detailed experimental methodology and testbed implementation are added in Section 4. Also, the performance comparison between SampleLite and RaCA is added in Section 4. These improvements have significantly validated the effectiveness of RaCA.
3. Design of RaCA
This section presents the design details of RaCA, a joint rate and channel adaptation scheme.
3.1. Overview of RaCA
As shown in Figure 2, RaCA consists of four key components: LSD (Link State Diagnoser), PTFD (Packet Transmission Failure Differentiation), TSRA (Two-Stage Rate Adaptation), and QCA (Quorum-based Channel Adaptation). LSD is primarily responsible for collecting current link conditions, which are gained at the receiver side. We design the LSD based on the information gained at the receiver side. PTFD diagnoses the exact causes for a packet transmission failure when it occurs. TSRA is responsible for identifying and setting transmission rate to the best one. QCA implements a channel adaptation system using cyclic quorum, where each transceiver pair can rendezvous on another different channel to continue communication. Based on the link condition information provided by LSD and the differentiating results from PTFD, RaCA will choose to activate TSRA or QCA to select best rate or channel to work on. TSRA will select an appropriate rate to achieve best throughput performance if the current link works very well or suffers severe channel-errors. QCA will choose another channel to work on if current channel suffers severe interferences. Next, we will give the design details of these components in RaCA.
3.2. PTFD and LSD
It is well known that a packet transmission failure in IEEE 802.11 networks can happen due to either interference or weak signal. It is very important to discern the exact cause of a packet transmission failure for rate adaptation problem, once it occurs. Meanwhile, the IEEE 802.11n standard provides more than 300 MCs, which is a large search space for rate adaptation scheme to select an appropriate rate. Therefore, it is crucial to reduce the search space for an efficient RA method. Based on the above two considerations, the goal of LSD has twofold. On the one hand, LSD provides enough information for PTFD to discern the exact causes of a packet transmission failure when it happens. Thus, the RaCA can decide to initiate TSRA or QCA to work. On the other hand, LSD must provide sufficient information for TSRA to reduce search space and select an appropriate rate efficiently.
PTFD. Some loss differentiation schemes are proposed in IEEE 802.11 networks [13, 28, 29]. The most recent research result implements a loss differentiation framework called BLMon [26] for high-speed IEEE 802.11n networks. BLMon uses frame aggregation and block acknowledgements to accurately differentiate losses with minimal overhead, in particular, to differentiate the loss causing collisions, hidden terminals, and noises. BLMon defines three metrics, namely, BII (Burst Isolation Index), ARET (number of A-MPDU (MAC Protocol Data Unit) RETries), and H-PER (History based PER, Packet Error Rate), based on the block acknowledgement information. The specific discerning algorithm can be referenced to the description in [26]. In our paper, PTFD adopts BLMon to provide exact causes of packet transmission failures for RaCA. PTFD does not differentiate the collision from the hidden node and takes both situations as interferences. Specifically, PTFD infers interferences that happened if the transmissions suffer distinct bursty losses as indicated by the metric BII. Otherwise, channel-errors are the main factor that leads to transmission failure.
LSD. Firstly, LSD uses RSSI and CSI information to accurately measure the current channel conditions. RSSI is a measure in the physical layer of the power observed at the antennas and is reported to MAC layer in the IEEE 802.11n. Therefore, the RSSI measurements are already available at the receiver side. While the CSI is a measure at the OFDM subcarrier level to support MIMO operation and the IEEE 802.11n, NICs report this information in a standard format. LSD is based on the following two insights. One is the monotonic relationship between the number of spatial streams and channel width of each IEEE 802.11n feature and the averaged RSSI of a link as presented in the SampleLite [2]. On the other hand, the CSI provides a much richer source of information than RSSI [19]. We also have conducted extensive experiments to analyze the characteristics of RSSI and CSI. In experiments, we use one transmitting antenna and two receiving antennas, 20MHz channel bandwidth. For CSI value, we randomly choose three subcarriers for every two datastreams. Figures 3 and 4 show the instant RSSI value and the instant CSI value for different spatial datastreams in a typical indoor laboratory environment. Figures 5 and 6 provide the mean square error for RSSI and CSI. As shown in figures, the change of instant CSI value is more obvious than that of RSSI, even for different subcarriers and datastreams. Although RSSI does not reflect the frequency selective attenuation characteristics of IEEE 802.11 channels, over a relatively long period of time, RSSI is more stable than CSI.
(a) Instant CSI value for datastream 1
(b) Instant CSI value for datastream 2
(a) Mean square error of CSI for datastream 1
(b) Mean square error of CSI for datastream 2
Based on the above findings, we choose RSSI to determine channel width, number of spatial streams, and guard interval, which can reduce the complexity of implementation for RaCA. And we use CSI to design an effective SNR for a multicarrier channel to determine the modulation coding schemes and other parameters. Therefore, we combine these two kinds of information together to achieve different goals. We use RSSI, a coarse-grained indicator, to decide which configure settings to search and use CSI to design an effective SNR to determine the modulation coding schemes and other parameters. The specific method to compute CSI can be referred to in [19]. However, due to signal attenuation and multipath effect, the signal at the receiver side cannot accurately reflect the real channel conditions at the transmitter side. Therefore, in our paper, the averaged RSSI is measured at the receiver side, while SampleLite measures RSSI at the transmitter side. Secondly, in order to reduce transmission overhead, RaCA disables the RTS (Request to Send)/CTS (Clear to Send) function and uses the block acknowledgement function provided by the IEEE 802.11n standard to differentiate the true causes of transmission failures. To achieve this target, LSD modified the block acknowledgement message, a 64-bit bitmap, and fed it back to the transmitter side.
Because both hardware and communication environments can affect the RSSI readings [30], an individual instant RSSI value cannot accurately reflect the true channel state. In this paper, we adopt the error-based filter [31] to design a conservative RSSI estimator which can deal with the RSSI fluctuation problem. The RSSI estimator is formulated as follows:where is the new estimation value of RSSI, is the old estimation value of RSSI, and is the actual measured value of RSSI. is a smoothing coefficient which is not constant. is calibrated by the following formula:where shows the predictive power for the error-based filter and controls the error deviation of the error-based filter. So [31] named as estimator error. When the error-based filter generates the estimations matching well with measurements, the larger weight is given to the old estimations by enlarging the value of smoothing factor . If not, it decreases the weight of the old estimations by reducing the value of smoothing factor . is the absolute difference between the old estimation and the actual measured value. In error-based filter, does not use the original difference directly. On the contrary, error-based filter balances the estimation error via exponentially weighted moving average filter: is the maximum estimation error recently measured. is assigned dynamically to minimize the estimation error in different experiment scenarios.
3.3. TSRA
TSRA tries to identify and set its transmission rate to the best one. Unlike other rate adaptation schemes, TSRA implements a two-stage rate adaptation scheme based on the information gathered by PTFD and LSD. IEEE 802.11n standard specifies different parameters for different MCs. For example, Table 1 shows the MCs parameters for 40 MHz channel with three spatial streams. In order to reduce the search space, TSRA decides the channel width, number of spatial streams, and guard interval based on the RSSI information in the first stage. In the second stage, TSRA chooses the specific rate for current channel based on the CSI information as shown in Table 2. The pseudocode for TSRA is presented in Algorithm 1. The definitions of variables used in Algorithm 1 are listed in Table 3. The values of , , and are determined by the specific simulation environment, and they are set to -45, -70, and -60 in our simulations. is set to -45 in mobile environment and -60 in static scenario. The value of is set to 90%. As shown in Section 4.2, these thresholds work well in different scenarios. The specific method to predict the PER using the CSI can be referred to in the descriptions in [32].
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||
3.4. QCA
If current link condition is under severe interferences, reducing the transmission rate will degrade the network performance dramatically. The most direct way is to use another channel to communicate. QCA implements a channel adaptation scheme based on the quorum system, where each transmitter and receiver pair can rendezvous on another different channel to work on. Next, we first review the background on quorum system. Then, we will present the design details of QCA.
3.4.1. Background on Quorum System
Next, we retrospect some background information on quorum system, which will be applied throughout the following section.
Definition 1. Given , a cycle period, let be a limited universal set having elements. A quorum system under is a set of nonempty subsets of and meets the following intersection feature:Each is a subset of , which is called a quorum.
Definition 2. A set modulo , is called a cyclic -difference set if for each (mod ) here are just ordered pairs , such that - (mod ).
Definition 3. A set modulo , is called a relaxed cyclic -difference set if for each (mod ) here exists at the very least one ordered pair , such that - (mod ).
Theorem 4. A collection of sets modulo , , is a collection of cyclic quorum sets when and only when is a relaxed cyclic (, )-difference set.
The authors in [33] have proved Theorem 4, which shows that we can build a cyclic quorum system by cyclic difference sets. Meanwhile, the authors in [34] have shown that any in under must have a cardinality . In other words, given , the minimal size of the quorums in cyclic quorum system is . In practice, the theory of lower band can be constructed by building Singer difference sets [35]. Next, the Singer difference sets theorem is presented, which is demonstrated by [36].
Theorem 5. Let be a prime power. Then, there is a -difference set under , which is called a Singer difference set. Here, , , and denotes a collection of nonnegative integer less than .
Because and , approximates to the lower bound .
Definition 6. Given an integer and a quorum in a quorum system under , we define .
Definition 7. A quorum system under is considered to have the rotation closure feature if , , .
The authors in [34] have proved the following theorem.
Theorem 8. The cyclic quorum system meets the rotation closure feature.
3.4.2. Design of QCA
Firstly, we construct a channel adaptation system using quorum system. Without loss of generality, we presume to build a channel adaptation system , where each transceiver pair can meet on different channels. An illustration of channel adaptation system system with is shown in Figure 7. In Figure 7, a channel adaptation system was built using the quorum system over the limited universal set . The elements in are the channels that the transceivers can work on. If the transceiver pair chooses one quorum in , then they can rendezvous on three channels.
Next, we introduce the algorithm to build the . Without loss of generality, we presume that each consists of fragments, and every fragment consists of timeslots. Therefore, the duration of every is . Particularly, we presume that and , and is the rendezvous channel set. Then, the creation step is as follows.(1)Firstly, we build a universal set and construct a under . Let . Then, the following quorums are built as follows:(2)We build using the following steps by the quorum .(i)We use the following formula to construct the channel adaptation schedule for the first fragment of timeslots: where is the channel that transceivers work on and is selected from the set at random. indicates the channel that the transceivers should work on in the timeslot.(ii)Repeat the above step to build channel adaptation schedule for other two fragments. We should pay attention to the rest of fragment; the timeslot index ought to be the modulo over for building the above formula .(3)For other quorums in , e.g., , , and , repeat Step to build other three channel adaptation schemes to . The four channel adaptation schedules, namely, are the elements in the collection of , which has a duration of .
The channel adaptation schedules in are demonstrated in Figure 8. We construct a channel adaptation schedule in using a quorum in . Therefore, we get . Note that , , there are two corresponding quorums used to build and , respectively. The period of every channel adaptation schedule in is , and . Owing to the intersection feature of , in each duration of , and rendezvous at least times on a particular channel . As shown in Figure 8, the transceiver pairs choose any two channel adaptation schedules in ; then they can rendezvous at least 3 times in each duration of . The pseudocode for QCA is presented in Algorithm 2.
| ||||||||||||||||||||||||||||||||||||
4. Performance Analysis
Next, we validate the performance of RaCA under a variety of network environments scenarios. To validate the efficiency and the feasibility, we have implemented RaCA in NS-3 (version 3.27) [37] and commercial 802.11 NIC.
We compare the performance of RaCA against Minstrel-HT and SampleLite in terms of aggregated throughput. The aggregated throughput is the sum of data rates achieved by all the mobile stations in the network. The Minstrel-HT [16] has been implemented in NS-3 [38]. The SampleLite is a latest research that provides substantial gain relative to Minstrel-HT in terms of goodput performance. We have implemented SampleLite and RaCA in NS-3 and testbed.
4.1. Simulation Setup
The network topology used in our simulation consists of a fixed AP and multiple mobile terminals. A point-to-point link collects the AP to a local area network. Figure 9 shows our 24-node simulation environment. We establish TCP and UDP datastreams from mobile terminals to local area network nodes. The traffic generator produces a constant traffic rate of 2 Mbps, and each data packet has 1024 bytes. The application layer sends a total of 1,200 packets. In the simulations, two mobile scenarios are considered, and the mobile speed of terminal is uniformly distributed between 1 and 10, 10 and 20, respectively. Also, we simulate two types of network density, 10-mobile stations and 24-mobile stations, to analyze the effect of interference level on rate adaptation schemes. All the simulation results are the average of 5 runs. The specific configuration parameters of simulations are listed in Table 4.
4.2. Simulation Results
Impact of Network Density (TCP). We first simulate two static network scenarios, 10-mobile stations and 24-mobile stations, to show the effect of interference on throughput performance. Figure 10 gives the simulation results. We can find that RaCA outperforms both SampleLite and Minstrel-HT in terms of aggregated throughput on average by 28%, 50.2% and 46.8%, 84%, respectively. Minstrel-HT performs worst in all scenarios because it reduces the number of spatial streams to lower bit rate if the channel is to be too lossy regardless of the causes leading to packet transmission failures. Particularly, in dense network scenario, Minstrel-HT aggravates the effect of interference due to longer transmission time and increases the contention level and likelihood of collisions. RaCA avoids this problem by adopting PTFD scheme to distinguish the loss pattern of packet transmissions and combining RSSI and CSI together to accurately and rapidly respond to MCs configurations. If the channel suffers severe interferences, RaCA chooses to work on another channel using QCA schemes. Although SampleLite achieves better performance compared to Minstrel-HT, it suffers significant performance degradation because it ignores the causes of transmission failures and reads the RSSI at the transmitter side. On the contrary, RaCA computes RSSI and CSI at the receiver side to deal with this problem.
Impact of Mobility (TCP). Figure 11 shows the aggregated throughput achieved by different mobility scenarios. As shown in the figure, both schemes suffer significant performance degradation when node moves faster. However, RaCA provides significant aggregated TCP throughput gain over SampleLite and Minstrel-HT in both environments, up to 7.1% and 47.5% in 1~10 speed scenario, 37.8% and 51.1% in 10~20 speed scenario, respectively. Minstrel-HT performs poorly because it reduces the number of spatial streams to lower bit rate quickly once transmission fails. This aggravates the impact of mobility as it takes more time to transmit packets and increases the likelihood of channel-errors due to node’s frequent mobility. Better performance with SampleLite in both scenarios is because it relies on RSSI to choose the MCs, which will provide maximum expected throughput for packet transmission. However, through analyzing the trace data, we found that SampleLite cannot respond to the most appropriate rate, which is because of incorrect and coarse-grained property of RSSI obtaining at the transmitter side. This is why RaCA outperforms SampleLite in all scenarios. RaCA can accurately and quickly respond to MCs search space and rate by combining RSSI and CSI information together, especially in high mobility scenarios.
Impact of Network Density (UDP). We also conduct experiments with UDP datastreams. Figure 12 shows that RaCA performs best in terms of UDP aggregated throughput compared to the other two schemes, up to 60.8% over Minstrel-HT and 28.5% over SampleLite in 10-mobile stations scenario, 74.9% over Minstrel-HT and 55.7% over SampleLite in 24-mobile stations scenario, respectively. Similar to the TCP scenario, RaCA achieves high gains in different network density scenarios. These gains are mainly attributed to the following two aspects. On the one hand, Minstrel-HT reduces the number of spatial streams to choose lower bit rate if the channel is to be too lossy regardless of the causes leading to packet transmission failures. In contrast, the TSRA algorithm provided by RaCA can probe best rate rapidly and accurately based on CSI and RSSI information. On the other hand, Minstrel-HT and SampleLite do not consider causes of the packet transmission failures, which will result in high packet transmission failures in intensive network scenarios, where channel communications suffer severe interferences.
Impact of Mobility (UDP). Figure 13 plots the UDP aggregated throughput measured in different mobile scenarios. We see that RaCA still works better than the other two schemes in both scenarios. In 110 speed scenario, we observe up to 51.5% and 17.4% higher aggregated throughput compared with Minstrel-HT and SampleLite. In 1020 speed scenario, RaCA also gives significant gains, which are up to 58.9% over Minstrel-HT and 45.2% over SampleLite. These gains are attributed to rapid response to channel condition changes. Particularly, in high speed scenarios, RaCA can converge to appropriate rate rapidly and accurately.
4.3. Testbed Implementation
Next, we introduce the testbed experiments conducted in a controlled laboratory environment. Figure 14 shows the floor plan of our testbed environments. Each point from L1 to L6 in Figure 14 represents a location of stations. The RaCA was implemented on a commercial IEEE 802.11n NIC with AR9380. We generate data traffic using Atheros CSI tool [39]. For each set of experiment scenarios, we send 200000 packets with 1000 bytes per packet at the application layer every 50 milliseconds and average the results of 10 runs.
Impact of Mobility. In this scenario, we generate datastreams from one station to an AP. When the station keeps static, RaCA achieves the best aggregated throughput by quick convergence to the right bit rate, as shown in Figure 15. SampleLite achieves much the same throughput performance with RaCA, while Minstrel-HT provides the worst performance. In mobile environment, the station keeps moving with walking speed. Thanks to the intelligent rate adaptation scheme, RaCA also provides the best throughput performance, achieving up to 50.1%, 70.4% higher throughput compared to Minstrel-HT and 20%, 30.7% compared to SampleLite in static and mobile scenarios, respectively. According to analyzing the data rates distribution measured in mobile environment, we found that RaCA dominantly uses the most stable rate compared to Minstrel-HT and SampleLite. It also demonstrates that RaCA can respond rapidly to channel state changes.
Impact of Interference. In this scenario, two static stations send data to AP simultaneously at the locations L1 and L2, L3 and L4, and L5 and L6, respectively. Also, we consider mobile scenarios, where three group mobile stations keep moving between L1 and L2, L3 and L4, and L5 and L6, respectively, and they send data to AP at the same time. In these scenarios, we keep station close enough to the AP to ensure that the AP does not receive packet incorrectly because of weak signal. Figures 16 and 17 show the aggregated throughput obtained by AP at different locations in static and mobile scenarios, respectively. RaCA shows significant throughput gains which are up to 72.2% over Minstrel-HT and 31.4% over SampleLite in static scenarios, 97.3% over Minstrel-HT and 47.9% over SampleLite in mobile scenarios. Thanks to the intelligent rate and channel adaptation scheme, RaCA can quickly switch to another channel when suffering from interferences. When mobile station moves away from AP, the station suffers from channel-errors. RaCA can fastly converge to right bit rate and achieve much better throughput performance than the other two protocols. In mobile station scenario, AP achieves relatively lower throughput for all three protocols.
5. Conclusions
In this paper, the rate adaptation problem for IEEE 802.11n networks was investigated, and the reason why it needs to adapt rate and channel together for intensive IEEE 802.11n networks was demonstrated. A joint rate and channel adaptation scheme termed RaCA was proposed and implemented using RSSI and CSI, which are collected and computed at the receiver side. The key idea is that RaCA chooses an appropriate rate using TSRA algorithm when the communication suffers severe channel-errors. Otherwise, RaCA chooses another channel to work on using QCA algorithm when the communication suffers severe interference. Simulation results demonstrate that RaCA outperforms Minstrel-HT and SampleLite in different simulation scenarios by up to 84% and 55.7% increase in aggregated throughput, respectively. Testbed implementations also show that RaCA gives significant gains over Minstrel-HT and SampleLite.
Data Availability
The data used to support the findings of this study are included within the article.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work is supported by the National Natural Science Foundation of China (NSFC) under Grants nos. 61872385, 61673396, 61772551, and 61801517 and the Fundamental Research Funds for the Central Universities under Grants nos. 18CX02133A, 18CX02134A, and 18CX02137A.