CHannel and Radio Interface Switching for Multi ... - Semantic Scholar

3 downloads 14248 Views 102KB Size Report
Abstract— Multi-radio Wireless Mesh Network (MWMN) has been recognized ... network to freely associate services and resources such as computing, storage,.
CHRIS: CHannel and Radio Interface Switching for Multi-Radio Wireless Mesh Network Chengchen Hu#*1, Kai Miao*2, John Vicente*3, Sanjay Rungta*4, Minjiao Ye*5 #

Dept. of Computer Science and Technology, Tsinghua University 1

[email protected]

*

Intel Corporation

2

[email protected] [email protected] 4 [email protected] 5 [email protected] 3

Abstract— Multi-radio Wireless Mesh Network (MWMN) has been recognized as a significant direction for the next generation wireless network. This paper presents a performance problem due to transmission oscillation that occurs in multi-radio wireless mesh network, which has not been reported by anyone in the research community. Such transmission oscillation introduces reduced throughput, greater network delay and unacceptable delay variations, which have all been identified in our experiments on a multi-radio test-bed. In this paper, to overcome this oscillation problem, we propose a mechanism which we call CHannel and Radio Interface Switching (CHRIS). In operation, CHRIS first carries out a metric comparison and then employs a two-phase switching process. This mechanism has been evaluated on a multi-radio test-bed and the results show that CHRIS is not only able to improve average performance, but also smoothes out variations in specific performance characteristics.

a type of ad hoc networking, WMN aims to diversify the capabilities of ad hoc networks [4]. Equipped with multiple radios in each node, Multi-radio Wireless Mesh Network (MWMN) can significantly mitigate interferences existing in the classical WMN environment, by transmitting over multiple radios simultaneously using orthogonal channels.

Routing and channel assignment (CA) have been widely investigated to leverage multi-radio multi-channel multi-hop wireless network in the literature. Routing determines the path and radio interface to be utilized, while CA selects the channel in a radio interface that should be used. A formal description presented in [5] optimizes the overall network throughput subject to fairness constraints on allocation of scarce wireless capacity among mobile clients. In [6], the authors proposed a distributed heuristic algorithm that produces near-optimal joint routing and CA solutions. The authors in [7] considered the I. INTRODUCTION interaction between channel assignment and distributed There has been increasing amount of interest from scheduling in multi-channel multi-radio WMNs and introduced academia and industry for Wireless Mesh Network (WMN), an approach to partition the network into sub-networks in and a growing number of services and applications are which simple distributed scheduling algorithms can achieve anticipated to run over wireless mesh infrastructure in the 100% throughput. Generally speaking, these algorithms firstly future. A P2P overlay network shares similar features with a pick up a channel and radio interface with the best metric wireless mesh network, such as no centralized control, under certain constraints (identification step). And secondly, it dynamic network topology, and localized operation [1]. We directly switches to a channel and radio interface identified envision a decentralized communication system [2], OverMesh, (switching step) in the identification step. The differences which is formed by wireless mesh networks and a highly among these algorithms are mostly in the first identification virtualized, converged computing and communications node- step, i.e., identify a channel and radio in terms of metrics, based architecture with emergent management capabilities. constraints, information exchange protocol and etc. But in OverMesh embeds the users inside the network to freely second phase, they all switch the channel and radio directly by associate services and resources such as computing, storage, default. Although former studies provide great insights into and bandwidth, therefore creating a network environment for identification of best channel and radio for WMN, they did not rapid network innovation. In our previous research, we have notice an important performance problem caused by the direct studied the overlay architecture and cross-layer search radio switching mechanisms. mechanisms in OverMesh [1] [2]. This paper will investigate In this paper, we will first show that performance can not the issues associated with multi-radio wireless mesh network. be automatically guaranteed by simply switching from one set Wireless Mesh Network is characteristic of low cost, easy of channel and radio to another set identified as better. In fact, deployment, and enhanced network capacity [3, 4]. Each node as shown in our experiments, without an effective switching in WMNs operates not only as a host but also as a router, mechanism, a MWMN system often cannot remain in a single forwarding packets on behalf of other nodes that may not be stable state as expected and oscillations in key performance within direct wireless transmission range of their destinations. characteristics would occur, which leads to lower average This type of multi-hopping offers increased reliability, throughput and larger average delay, as well as unacceptable coverage and reduced equipment costs over their single-hop network jitter. We will then provide some analysis based on counterpart, Wireless Local Area Network [5]. Instead of being our observations of the phenomenon described above, describe

Mesh connection is a virtual network connection in layer two. Please note that, only one IP address can be assigned to a connection. In MROMP, each mesh network has only one virtualized mesh connection, which means each mesh can be assigned with an IP address. For one mesh with multiple radios, the radios are not exposed to applications; however, the routing protocol is aware of the radios. Mesh Kernel Service provides necessary functions which can be used by underneath radios device drivers as well as mesh user space services, e.g., creation/destroy of mesh network, data transmission/receiving, radio register/unregister. Neighbor table is maintained in kernel and provides capability for user space routing protocol to set neighbor information. It also provides extensibility for Mesh Kernel Service to use neighbor information in different routing policy. Route table is maintained in kernel and is used for routing (radio interface and channel) lookup.

Fig. 1. Architecture of multi-radio OverMesh platform the root cause of the problem, and propose a solution in the form of switching mechanism, which we call CHannel and Radio Interface Switching or CHRIS. We will demonstrate by results from experiments that CHRIS is able to greatly increase performance in terms of throughput, network delay and jitter. The remaining of the paper is organized as follows. Section II presents the performance problem caused by the simple switching mechanism and the factors which affect the performance. Section III describes the proposed channel and radio interface switching mechanism in details and Section IV performs experiments to evaluate CHRIS. Finally in Section V, we conclude the paper. II. PROBLEM STATEMENT A. Multi-radio OverMesh platform This paper is a part of OverMesh project, and performs experiments from multi-radio OverMesh platform. The nodes in the platform participate in multi-hop wireless networking and can be either stationary or mobile. Please note that, each node equips with two Network Interface Cards (NICs). One provides 802.11a radio (using a NETGEAR 802.11a card), and the other provides 802.11g radio (using a Intel PRO/Wireless Network Connection). The two radios coexist in each node and work on orthogonal channels. The platform instantiates a prestandard IEEE 802.11s mesh network for each radio, which provides complete support of metric-based multi-hop routing at link layer, neighbor discovery, link quality measurement, and a user interface. The schematic architecture of the Multi-Radio OverMesh Platform (MROMP) in each node is shown in Fig. 1. Radio is the wireless radio which provides real network communication. For each radio, it has its own device driver to provide necessary data and call functions to register to MROMP. In such way, MROMP can utilize the proper radios for mesh communication.

Mesh user space services provide a set of APIs to connect the mesh kernel and routing protocol modules. Those services includes: interfaces to define routing protocol related packet type, interfaces to transmit and receive protocol related packets, interfaces to configure kernel route table and neighbor, mechanism to receive kernel events. Routing management enables routing related functions. In our implementation, the routing protocol is AODV [8] and the metric is ETX [9] (expected transmission count). Any routing protocol (e.g., DSR [10]) can be implemented in the routing protocol module which sends and processes routing, and any metric (e.g., WCETT [11]) measurement function can be deployed in metric management module. B. Experiment configurations and results There are three nodes, A, B and C in the experiment, each of which equips with two radios: 802.11a and 802.11g. To simply but clearly illustrate the problem, channel 36 is configured if 802.11a radio is selected and channel 1 is utilized if 802.11g radio is selected. A UDP flow sends packets with a constant rate of 27Mbps from A destined for C. A and C can not communicate with directly and B is utilized as a relay node forwarding packets on behalf of A to C. In the experiment, the metric used to evaluate path quality is ETX [9]. Let M 1 and M 2 denote the metrics for two radios. Without lose of generality, in this paper, we suppose that M 1 > M 2 means metric of radio one is better than the one of radio two, and vice versa. The traditional MR mesh driver uses simple switching algorithm as follows: if M 1 > M 2 switch to radio one; else switch to radio two; We use iperf to collect performance data in the experiment and show the performance in Fig. 2. The experiments using

Throughput (Mbit/s)

(a)Throughput 30 25 20 15 10

0

50

100

150 time t (s) (b) Delay

200

250

300

Delay (ms)

4

Study from the simple switching method, we find out that two factors lead to such oscillation. The first cause is the aggressive metric comparison method and the other cause is the undesired global synchronization.

3 2 1 0

0

50

100

150 time t (s) (c)Loss

200

250

300

0

50

100

150 time t (s)

200

250

300

Loss (%)

60 40 20 0

Fig. 2. Performance curves of the experiment other metrics (e.g., WCETT [11]), is omitted for brevity, which display very similar trends. Fig. 2 (a) is the received data throughput in node C, which shows that the throughput varies between alternate extremes, wherein one extreme is around 27 Mbits/s and the other counterpart is about 16 Mbits/s. When the two hops in this experiment scenario employ different radios, the throughput is 27 Mbits/s as the flow being sent. However, the throughput is reduced to half of the actual maximum transmission rate1, 16 Mbits/s, if the two hops using the same radio to communicate, because of the radio interference. In Fig. 2 (b), we observe that the delay jitter is quit large. The maximal delay could be ten times of the minimal one and it is not acceptable for delay sensitive applications, e.g., VoIP applications. Periods with 40% loss are also demonstrated in Fig. 2 (b) when the throughput reduced to 16 Mbits/s. The oscillations reduce the expected performance of multi-radio wireless mesh, and more unfortunately, there is no sense to indicate a stop of such performance oscillations. C. Causes analysis In order to communicate with each other, it is required that both nodes in each link use the same channel on the same radio. Thus the experiment scenario must be in the one of four states as illustrated in Fig. 3: (1) State 1 denoted as (0, 0): link AB and link BC both use channel 36 on 802.11a; (2) State 2 denoted as (0, 1): link AB and link BC select channel 36 on 802.11a and channel 1

1 on 802.11g, respectively; (3) State 3 denoted as (1, 0): link AB uses channel 1 on 802.11g, and link BC selects channel 36 on 802.11a; (4) State 4 denoted as (1, 1): link AB and link BC both work on channel 1 of 802.11g. Obviously, State 2 or 3 is the state we expected because it has least interference from hidden terminals and achieves the lowest collision probability. The throughput is 27 Mbits/s in these two states as shown in Fig. 2. But in fact, the system oscillates among the four states, and thus affects the performance as we demonstrated in Fig. 2.

The maximum transmission rate of 802.11a or 802.11g card is only 32Mbits/s (not 54 Mbits/s) for application data as shown in our experiments. It is caused by the overhead for MAC control, signaling, preamble and etc.

In simple switching method, a switching operation will be performed, if the metric of the other radio/channel is better than the one currently used. The metric comparison method is too aggressive and does not consider the metric changes by channel switching. An example is given to state the problem. Suppose in the experiment mentioned above, the system is currently in state 2, i.e., the link AB uses channel 36 on 802.11a, and link BC uses channel 1 on 802.11g. Since there is no traffic transmitting using 802.11g over link AB, the metric (denoted as M g ) of 802.11g should be better than the one of 802.11a (denoted as M a ) and a switching to 802.11g will be triggered. After the switching, the metrics will be changed and 802.11a is better this time because it is free of transmission.) and make it jump back to 802.11g. Since only one flow exists in this case, no matter which interface is selected in each link, the node could find the other radio is providing better service and trigger a switching to use that radio interface card. Therefore, each node uses two radios alternately and can not stay in the status we expected. Considering the scalability and deployment, the distributed control methods are always preferred. In the distributed manner, each node probes the quality of the links to its neighbor independently to identify a better radio interface/channel if exist. Then the node intelligence will conduct a switching operation to the identified better radio/channel. In such cases, each node may switch to the same best radio or channel (depends on the metric) and then soon worsen quality of the radio/channel. For example, we detect state transfer from state 1 to state 4 in the experiment. III. CHRIS: CHANNEL AND RADIO INTERFACE SWITCHING Having the above analysis, we propose CHRIS to solve the problem, which consists of two parts: a metric comparison part and a two-phase switching part. Metric comparison part is to compensate for the original aggressive metric comparison method; while the two-phase switching part is introduced to avoid undesired global synchronization. A. Metric comparison Suppose each node has two radios. In time t, radio 1 is in use with metric M 1 and the metric of the candidate radio 2 for potential switching is M 2 . Ideal metric comparison logic should take into account the metric changes caused by the

Throughput (Mbit/s)

(a)Throughput 28 26 24 22

0

50

100

150 time t (s) (b) Delay

200

250

300

0

50

100

150 time t (s) (c)Loss

200

250

300

0

50

100

150 time t (s)

200

250

300

Delay (ms)

1.5 1 0.5 0

Fig. 3 system states of experiment one switching, and compares M 1 with M 2' = M 2 + a , instead of comparing M 1 with M 2 (a is the metric difference of the radio 2 after switching). It depends on the calculation of the specific metric to estimate a, however, it is difficult to find a general estimation method for different metrics since the calculations of metrics vary a lot. To avoid the estimation of a and to make our scheme be a general one, we modify the metric comparison based on the ideal one. M1 ∧ M 2 + a M1 / M 2 ∧1+ a / M 2 M 1 / M 2 ∧ TH

{ Δt = random(1, mt );

0

Fig. 4. Performance curves of CHRIS using two radios

if

(2)

M 1 (t + Δt ) < TH M 2 (t + Δt )

switch to radio 2; (3)

B. Two-phase switching mechanism To avoid the undesired global synchronization, we further introduce a two-phase switching mechanism. The switching to a new radio will not be triggered immediately after one comparison of M 1 / M 2 and TH. If M 1 / M 2 < TH , one more comparison will be performed after a random time interval (in the unit of an switching time slot). Only when it holds true that M 1 / M 2 < TH in both two comparisons, the switching will be finally executed. Combined with the metric modification and the two-phase switching, the pseudo-code CHRIS algorithm is as follows, where mt is the maximum delay interval.

M 1 (t ) < TH M 2 (t )

5

(1)

We compare M 1 / M 2 with a pre-defined threshold TH (TH > 1) and M 1 / M 2 < TH means a stable and better status could be achieve to switch to radio 2.

if

Loss (%)

10

} C. Parameters setting In CHRIS, there are two parameters: TH and mt. TH should be larger than one as we show before. If the TH is set as a value larger than two, the system can be stably select radios as we will demonstrated in the next section. However, TH should be not too large, since large TH may lead to slow reaction of the system. In the evaluation, we observe that small mt (smaller than 4) does not solve the problem and large mt (larger than 100) slow the system reaction time. Any value between 5 to 100 should be an acceptable choice. In the evlauation Section IV, the TH and mt are configured as four and ten respectively. IV. EVALUATION The evaluation of CHRIS is performed on the aforementioned multi-radio OverMesh platform. To compare with the experiment result in Section B, the experiment environment is configured similarly. Three nodes, A, B and C are employed, each of which equips with two radios: 802.11a and 802.11g. A UDP flow sends packets with constant rate of 27 Mbps from A destined for C with the relay of B. The performance curves of CHRIS are illustrated in Fig. 4. We observe that, these curves are much more stable compared with the ones in Fig. 2. The bandwidth is about 27 Mbps,

(a)Throughput Throughput (Mbit/s)

18 16 14 12 10

0

50

100

150 time t (s) (b) Delay

200

250

300

Delay (ms)

2 1.5 1 0.5 0

0

50

100

150 time t (s) (c)Loss

200

250

300

0

50

100

150 time t (s)

200

250

300

80 Loss (%)

performance), the average performance measures have been greatly improved. The average performance and performance variation is illustrated in Table 1. The average performance results obtained by the experiments when two radios are utilized are better than the single radio scenario. Compared with the simple switching methods in two radios scenario, the proposed CHRIS improves the throughput by about 12.6%, provides a 2.39 times shorter mean delay, and decreases the loss rate from 13.83% to 0.88%. Besides, we observe that, CHRIS diminishes the performance variations in simple switching to a quit low level.

60 40 20

Fig. 5. Performance when using only one radio which is almost the same as the rate we send packets. Meanwhile, the delay and loss is quit low with little variation. The reason is that it can stay in the State 2 or State 3 now with the help of proposed CHRIS. The results are not sensitive with the variation of parameter TH. Note that, the performance is also relatively stable using only one radio. But the bandwidth is almost halved, the delay and the loss are tremendously increased. The results are revealed in Fig. 5. This can be demonstrated, when we only activate a same radio in each node.

V. CONCLUSIONS The performance oscillation problem for multi-radio wireless mesh network is addressed in this paper. With a multi-radio OverMesh platform, we demonstrated and analyzed performance variations caused by metric comparison operations and undesired global synchronization in existing algorithms. To better utilize a multi-radio wireless mesh network or platforms, a new mechanism, CHRIS, is proposed for improving the average performance and reducing the performance variations. Our experiments show that CHRIS has obvious advantages in terms of both average performance and performance variation. Our future work in this area is to study the proposed mechanism with more sophisticated experiments in a large scale network with various topological characteristics. REFERENCES [1]

[2]

[3]

[4]

As we eliminate the oscillations in States (as well as the [5]

Table 1. Performance comparisons Single radio

Two radios, Simple Switching

Two radios, CHRIS

Mean Throughput

15.91 Mbps

23.77 Mbps

26.78 Mbps

Mean Delay

0.72 ms

0.42 ms

0.176ms

Mean loss

40.93%

13.83%

0.88%

Variation of Throughput

0.26

25.20

0.046

Variation of Delay

0.05

Variation of Loss

0.0007

[6]

[7]

[8]

[9] 0.211

0.011 [10]

3.43

0.0001 [11]

G. Ding, J. Vicente, S. Rungta, D. Krishnaswamy, W. Chan, and K. Miao, "Overlays on Wireless Mesh Networks: Implementation and Cross-Layer Searching," presented at International Conference on Management of Multimedia Networks and Services, Dublin, Ireland, 2006. J. Vicente, S. Rungta, D. Gang, D. Krishnaswamy, W. Chan, and M. Kai, "OverMesh: network-centric computing," Communications Magazine, IEEE, vol. 45, pp. 126, 2007. A. Raniwala and T.-c. Chiueh, "Architecture and algorithms for an IEEE 802.11-based multi-channel wireless mesh network," presented at IEEE INFOCOM 2005, 2005. I. F. Akyildiz, X. Wang, and W. Wang, "Wireless mesh networks: a survey," Computer Network, vol. 47, pp. 445-487, 2005. M. Alicherry, R. Bhatia, and L. E. Li, "Joint Channel Assignment and Routing for Throughput Optimization in Multiradio Wireless Mesh Networks," Selected Areas in Communications, IEEE Journal on, vol. 24, pp. 1960, 2006. H. Wu, F. Yang, K. Tan, J. Chen, Q. Zhang, and Z. Zhang, "Distributed Channel Assignment and Routing in Multiradio Multichannel Multihop Wireless Networks," Selected Areas in Communications, IEEE Journal on, vol. 24, pp. 1972, 2006. A. Brzezinski, G. Zussman, and E. Modiano, "Enabling Distributed Throughput Maximization in Wireless Mesh Networks - A Partitioning Approach," presented at ACM MOBICOM 2006, Los Angeles, CA, 2006. C. E. Perkins, E. M. Belding-Royer, and S. Das, "Ad hoc On-Demand Distance Vector Routing," presented at 2nd IEEE Workshop on Mobile Computing Systems and Applications, New Orleans, LA, USA, 1999. D. S. J. D. Couto, D. Aguayo, J. Bicket, and R. Morris, "A highthroughput path metric for multi-hop wireless routing," presented at ACM MobiCom 2003, San Diego, California, USA, 2003. D. B. Johnson, D. A. Maltz, and J. Broch, "DSR: The Dynamic Source Routing Protocol for Multi-Hop Wireless Ad Hoc Networks," in Ad Hoc Networking: Addison-Wesley, 2001, pp. 139-172. R. Draves, J. Padhye, and B. Zill, "Routing in multi-radio multi-hop wireless mesh networks," presented at ACM MobiCom 2004, 2004.

Suggest Documents