Distributed MAC protocols and priority oriented ... - Wiley Online Library

INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMS Int. J. Commun. Syst. 2003; 16:475–495 (DOI: 10.1002/dac.600)

Distributed MAC protocols and priority oriented scheduling for a PLC access network S. Sundaresan, S. Anand, S. Srikanthn,y and C. N. Krishnan AU-KBC Research Center Anna University, M.I.T. Campus, Chennai 600044, India

SUMMARY In this paper, we propose a distributed channel allocation scheme for the uplink data traffic and implement a priority oriented scheduling for the downlink data traffic for a powerline access type network. We study the performance of the uplink and downlink protocols in terms of the buffering probability, mean delay, statistical dropping probability and the number of subscriber stations supported. The distributed channel allocation scheme uses a modified carrier sensing mechanism and a centralized collision resolution scheme. We model the channel occupancy of uplink data calls using a continuous time Markov chain (CTMC). We derive the expressions for mean delay, statistical dropping probability, and utilization by solving the CTMC. We also perform extensive simulations to complement our analytical results. Our results indicate that, the distributed channel allocation approach achieves 10% improvement in spectrum utilization compared to a centralized reservation protocol. We also implement a priority oriented scheduling policy for the downlink data traffic. In the system we consider, the downlink channels are adaptively modulated based on their signal strength. We maximize the instantaneous link utilization by using MaxFlow algorithm. Simulation results for downlink scheduling show that the mean buffering delay for the real-time variable bit rate traffic is zero and that the system supports 170 access stations. We also show that system with adaptive modulation achieves 26% savings in transmitted power compared to the system without adaptive modulation. Copyright # 2003 John Wiley & Sons, Ltd. KEY WORDS:

PLC; OFDM; DMAC; CTMC; scheduling; MaxFlow; adaptive modulation

1. INTRODUCTION The tremendous growth in telecom and information technology sectors has nurtured the need to provide high bit-rate services to the customers at low costs. The wired communication technologies like cables, optical fibres, hybrid fibre coax (HFC) provide high quality services, but at the cost of high installation expenses. As these expenses increment per length of the wire, their feasibility as ‘last mile access’ network remains to be seen in developing countries like India. In rural areas, where the subscriber density is much lower, installing new cables for communications would result in high ‘cost per subscriber’ ratio. The PLC

n

y

Correspondence to: S. Srikanth, AU-KBC Research Center Anna University, M.I.T. Campus, Chennai 600044, India. E-mail: [email protected]

Published online 9 May 2003 Copyright # 2003 John Wiley & Sons, Ltd.

Received 1 November 2002 Revised 15 January 2003 Accepted 19 February 2003

476

S. SUNDARESAN ET AL.

access network uses the existing low voltage(LV-220/440V) powerline segment to carry information between subscribers and a central base-station. As the powerline infrastructure is already present, implementing a powerline access network is relatively cheaper. The powerline medium is characterized by attenuation, multipath delay spread and various types of noise [1, 2]. Various electrical loads such as mixers, motors, lamps connect to the powerlines resulting in varying line impedances. This causes the signals to reflect producing multipaths and severe inter-symbol-interference (ISI). In order to combat this ISI, high complexity equalizer is needed in single carrier transmission techniques. Impulse noise in powerlines cause burst errors and render the medium unusable for a short duration of time. It causes high BER when single carrier transmission techniques are used. It can be seen that the powerlines are quite similar to wireless with respect to the medium characteristics. Studies in References [3] and [4] reveal that orthogonal frequency division multiplexing (OFDM) is an ideal physical layer modulation technique for such time variant media like wireless and powerlines. OFDM is a multi-carrier technique and the data is transmitted parallelly over several sub-carriers. Longer symbol time duration and cyclic prefix of OFDM symbols combats ISI caused by multipath delay spread and also minimizes the effect of burst errors [5, 6]. Hence, we consider OFDM as the physical layer modulation based on which we study the Layer2 protocols. The advances made in the physical layer should be complemented by efficient MAC and scheduling policies in order to realize a viable PLC system. Hence, we propose a novel distributed MAC(DMAC) for uplink traffic and a priority scheduling policy for the downlink traffic for an OFDM based PLC access network. A token passing MAC protocol for powerline LAN based on a spread spectrum physical layer has been studied and performance metrics such as throughput and utilization have been computed in Reference [7]. A carrier sense CDMA protocol based on overload sensing and packet admission control has been studied in Reference [8]. In this work, carrier sensing is used to determine the traffic load in the system based on which the length of the spread codes used to modulate the data is chosen. Packet admission control has also been implemented to prevent collisions that are inherent in a CSMA network. Though the study is aimed for LANs, the concepts such as carrier sensing and variable spreading can be used in access type networks also. Media access protocols for PLC access type network with OFDM physical layer has been studied in References [9] and [4]. In Reference [9], a centralized reservation protocol was evaluated for voice and data traffic. The data connections were considered to be circuit switched and blocking probability for voice and data connections computed. Moreover, uplink and downlink data traffic are considered as symmetrical. Data traffic is known to be bursty and also the traffic is biased more towards downlink. Hence, the assumptions considered in Reference [9] about data connections would not be applicable. In Reference [4], a reservation MAC protocol was studied for uplink data traffic alone. Downlink scheduling has not been considered in References [9] and [4]. Wireless protocols has been adopted for powerlines in References [10] and [11]. Though powerline medium is similar to wireless with respect to multipath delays and attenuation, it does not suffer from multiple access interference which is the main bottleneck in wireless. Hence, the protocols studied for wireless may not entirely suit PLC protocols. We focus on using the existing powerline infrastructure as access type networks in a residential scenario. We consider OFDM/FDMA as the multiple access scheme for our study as in Reference [9]. Frequency division duplexing(FDD) is considered for uplink and downlink Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

477

DISTRIBUTED MAC PROTOCOLS

traffic. The FDD spectrum is shown in Figure 1. An OFDM sub-carrier or a cluster of OFDM sub-carriers form a sub-channel.z The considered PLC network supports voice, data and video services. Since the distributed channel allocation approach is not applicable to voice services, we do not study it’s performance. But we reserve channels for voice calls for the sake of it’s coexistence in any network. For data traffic, we compute buffering probability, mean delay, statistical dropping probability and the number of stations supported. We outline the organization of the paper further. In Section 2, we describe the fundamentals of the considered PLC system. In Section 3, we propose a distributed MAC protocol based on carrier sensing mechanism for uplink traffic. We perform Markovian analysis that models the distributed MAC. In Section 4, we propose a downlink priority scheduling. The scheduling is done with adaptive modulation of OFDM sub-channels. We combine the priority scheduling with a graph theoretical approach called MaxFlow, that maximizes instantaneous link utilization. Finally we conclude highlighting our contributions in this paper.

2. SYSTEM MODEL A PLC access network consists of a central base-station (BS) and access units ðAUs Þ:} The BS acts as the gateway between the backbone network (telephone trunks, ATM, DSL, etc.) and the PLC access network. The AUs are situated at the customer premises. The AUs consist of a PLC modem with appropriate interfaces to terminal equipment like telephones, personal computers, etc. The network topology of the PLC access network is a logical bus structure as shown in Figure 2. All the system model parameters that we have considered in this work such as bandwidth, guard time, coding rate, etc., are obtained from Reference [9] and is shown in Table I. We have assumed that the data losses due to channel related impairments are taken care by physical layer techniques and link layer retransmissions. We have not focused on link layer retransmissions, because as far as MAC is considered, it does no distinction between original or retransmitted frame. We also assume that, hidden node problem does not exist in powerlines

VOICE

UPLINK

SIGNALLING & RESERVED CHANNELS

DATA

DOWNLINK

34

UPLINK

68

DOWNLINK

84

120

128 SPECTRUM

Figure 1. Frequency division duplexing spectrum used for in our system. z }

Subchannel and channel will be used interchangeably unless or otherwise specified. Access units and stations will be used interchangeably unless and otherwise specified.

Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

478

S. SUNDARESAN ET AL.

AU #2 BACKBONE NETWORK AC LINES 220/440 V BS

BS

- Base station

AU #1

AU #K

th

AU#K - K

Access Unit

Figure 2. Logical bus structure of powerline access network.

Table I. System model parameters. System parameter Bandwidth (BW) TGuard TOFDM Subcarrier spacing, df Number of subcarriers, N Coding rate, r Subchannel modulation Bitrate per sub-carrier

Value 5 MHz 2:8 ms 28 ms 39:0625 kHz 126 0.5 QPSK 32 kbps

[10]. In case hidden nodes are present in powerlines, Mangold et al. [11] suggests the use of powerline repeaters to combat the effect of hidden nodes. We reserve 6 sub-carriers for signaling and for replacement of active sub-carriers that are rendered unusable for a long time because of narrowband disturbances. Hence, 120 OFDM subcarriers are used for communication over powerlines. We reserve 68 sub-carriers for uplink and downlink voice traffic. Every voice connection is given a pair of sub-carriers and is held for the duration of the call. We reserve 16 sub-carriers for uplink data traffic and 36 sub-carriers for downlink data traffic. Every packet call either uplink or downlink, is assigned a pair of subcarriers resulting in a sub-channel of 64 kbps:

3. DISTRIBUTED CHANNEL ALLOCATION FOR UPLINK DATA TRAFFIC The existing literature on uplink MAC for PLC access networks have considered data connections as circuit switched and have assigned equal bandwidth to uplink and downlink data traffic [9]. This approach is inefficient as Internet data is bursty as shown in Figure 3 and the traffic load is strongly biased towards downlink [12]. Moreover, assigning a channel for an entire data session in time variant medium such as powerlines increases the bit error rate(BER), because the assigned channel may not retain it’s quality over the entire session. Hence, we assign channels on a packet by packet basis and it is known as dynamic packet assignment (DPA) [3]. We implement the DPA using a distributed channel allocation approach. We then derive a Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

479

DISTRIBUTED MAC PROTOCOLS

Data Session AU#2 Time Data Burst

Off Period

AU#3

Time

Figure 3. Timing diagram for uplink data traffic.

simple analytical model for the proposed scheme using Markov chains and compute access delay, statistical packet dropping probability and utilization. We also perform extensive simulations to validate our model. Finally, we compare the results with that of a centralized reservation MAC and bring out the advantages of DMAC. We define the performance metrics that are considered for uplink below. *

*

*

Access delay: The time interval between the instant when a packet is generated at an AU to the instant when it begins transmission. Statistical dropping probability: The percentage of packets whose access delay exceeds a pre-defined threshold. Example: If 2 out of 100 packets exceeds a pre-defined threshold of tmax ; then the statistical dropping probability is 2%: Utilization: The ratio of offered traffic load subject to statistical QoS constraints to maximum channel capacity.

3.1. Uplink data traffic model The data traffic in uplink mainly consists of packets generated from applications such as SMTP, TELNET, WWW requests for downloads and online interactive forms. These packets are not bulkier and do not have priorities between them. Every AU generates packet calls in a poissonian manner and can support more than one packet call at a time. We assume the same traffic model as used in Reference [4]. Each AU generates packets with mean rate of 0.208 packets per second. Each packet has an exponentially distributed size of mean 1500 bytes. When transmitted over a 64 kbps channel, a packet of mean size 1500 bytes holds the channel for 187:5 ms: 3.2. CSMA with centralized collision avoidance We describe the carrier sense multiple access protocol with centralized collision avoidance (CSMA/CCA) for the uplink data traffic in this section. In this protocol, the AUs selects the uplink channels for transmission by carrier sensing. It is much easier to implement carrier sensing in OFDM systems as the state of all sub-channels can be obtained simultaneously [13]. In systems with FDD, carrier sensing is implemented by switching to uplink reception mode and performing an FFT operation. It is intuitive that 1-persistent CSMA (i.e. continuous sensing) on uplink channels is not possible as it would block downlink reception. Hence, we present a Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

480

S. SUNDARESAN ET AL.

modified approach to CSMA, in which the AUs switch to uplink sensing, only when it has the knowledge that there is at least one uplink sub-channel free. The process of acquiring this knowledge and collision avoidance is described further. The CSMA/CCA scheme involves three main components. 1. A counter is maintained by the BS and it’s value is broadcast in the downlink signaling channel (Counter Broadcast Channel-CBCH). 2. All AUs have a variable register (VR) in which the current CBCH value from the BS is stored. 3. The AUs creates a fixed register (FR) for each packet it buffers. For example, if an AU buffers two packets, it creates two FRs and associates them to the two buffered packets. Figure 4 illustrates the channel access procedure between the BS and the AUs. The leftside of the timing diagram depicts the uplink channel state as measured by the BS and the counter value that is broadcast by the BS. The right side shows the value of VR in all AUs which is the current CBCH value from the BS and the action that an AU does when it generates a packet. As shown in stage (1) of Figure 4, the BS broadcasts a counter value of zero ðCBCH ¼ 0Þ when it senses free uplink channels. This results in all AUs to have a value of zero in their VRs. When a packet is generated at an AU, the value of VR is initially checked. If the AU finds a zero value in it’s VR, the AU switches to the uplink sensing mode and chooses a free channel. The packet is transmitted immediately on the free channel. As it can be seen, the only delay involved in this operation is the processing delay which is neglected. AU

BS Uplink channels available for transmission (1)

CBCH=0

Packet transmitted in one of the free channels

VR=0 indicates to AUs that channels are available for transmission Packet generated

Uplink channel state CBCH=1 change from "atleast one channel free" to "all channels busy" (2) AU sends '1' bit signalling to Upon receiving signalling BS from any AU, the BS CBCH=2 increments the counter.

VR=1 indicates to AUs that packets generated thereafter are to be buffered. Packet generated. The pakcet is buffered and FR=1 is mapped to the packet.

Uplink channel state change from "all busy" to "one free". Decrement counter

VR=1. Decrement in VR value causes AUs to decrement their FRs.

(3)

CBCH=1

VR=2

Buffered packet with FR=0 is transmitted

Figure 4. Channel access procedure for uplink data traffic. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

DISTRIBUTED MAC PROTOCOLS

481

The BS increments the CBCH value from zero to one as soon as it senses all channels to be busy as shown in stage (2) of Figure 4. This causes the VRs of all AUs to be 1 indicating that packets generated thereafter are to be buffered. Illustration: Consider an AU#1, which has generated a packet at this stage. The AU#1 checks the value of VR and on having found a non-zero value stored in it, the generated packet is buffered. The AU#1 then creates a FR with the current value of VR(=1) and associates it with the buffered packet. The AU#1 transmits an acknowledgment (ACK) bit ‘1’ to the BS. This is the only signaling that is done in uplink. Upon receipt of the ACK bit, the BS increments the CBCH value as shown in Figure 4. If another AU#2 generates a packet before a channel is released by a departing call, the process of stage (2) of Figure 4 results in AU#2 buffering the packet with FR=2 and the current value of CBCH and VR having a value of 3. The stage (3) of Figure 4 shows that the BS decrements it’s CBCH value when it senses a free channel due to a departing call. This causes the VRs of all AUs to decrement by 1. Illustration: When the AU#1 and AU#2 find that the current value of VR is less than the previous value, they decrement the FRs associated with the buffered packets. This results in the FR of AU#1 to be zero and that of AU#2 to be one. The AU#1 whose FR is zero switches to uplink mode and seizes the free channel. The AU#2 restrains from seizing the free channel and waits for it’s FR to become zero. From the above protocol it can be noted that collisions that are inherent in a CSMA scheme has been completely removed and also that the distributed way of channel access is maintained. We next describe the Markovian model for the DMAC and derive expressions for the performance metrics. The results of analysis are then compared with simulation results to validate the Markovian model. 3.3. Performance analysis of distributed MAC The uplink data traffic model follows a Poissonian distribution. Hence we develop an analytical model based on Markov chains for the CSMA/CCA strategy described above. Analytical expressions for queuing delay, statistical probability of dropping and utilization are derived. The analysis is done for the case when packets are not dropped from the queue. We do not analyse CMAC since the objective of this work is only DMAC and we had simulated CMAC for the sake of comparison with our distributed scheme. Delay distribution of packets in buffer is derived using which statistical dropping probability is computed. It is intuitive from our distributed scheme, that there is no priority assigned among the uplink packets and the packets undergo service in the order in which they arrive into the system. This means that packets are served on a FCFS basis. *

* *

The packets are assumed to arrive in Poissonian manner with rate kd ¼ lnd * M; where knd is the arrival rate per AU and M is the number of AUs. The holding times are assumed to be exponentially distributed with mean 1=md : We assume the number of packets in the entire system to be Nd (inclusive of those in transmission as well as those buffered)

The process Nd can be modelled as the queue length process of an M=M=nd queue where nd is the number of sub-channels allotted to uplink data traffic [14]. The channel occupancy of the uplink packets is modelled as a continuous time Markov chain (CTMC) and is shown in Figure 5. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

482

S. SUNDARESAN ET AL.

λd 1

λd

λd

2 µd

λd

λd nd

3 3µ d

2µ d

ndµ d

λd nd+1

ndµ d

ndµ d

Figure 5. CTMC model for the steady state process,Nd with a distributed MAC.

Writing global balance equations, ld pk
1 ¼ kmd pk ;

k4nd

ð1Þ

ld pk
1 ¼ nd md pk ;

k > nd

ð2Þ

where pk is the steady probability that there are k packets in the system. These equations can be solved to obtain, 8 k < p0 ðndk!rd Þ k4nd pk ¼ PrfNd ¼ kg ¼ ð3Þ : nndd rkd p0 nd ! k > nd P where rd ¼ ld =nd md : By applying the condition 1 k¼0 pk ¼ 1; we obtain the value of p0 as, " #
1 nX
1 1 k d X ðnd rd Þ ðnd rd Þn þ ð4Þ p0 ¼ k! n !nn
nd n¼nd d d k¼0 It is intuitive to observe that the probability of queuing is the probability that the number of packets in the system Nd is greater or equal to the number of channels available. Hence, 1 X pq ¼ PrfNd 5nd g ¼ pk ð5Þ k¼nd

Solving Equations (3)–(5), we can compute the probability of a call being buffered pq as p0 ðnd rd Þnd pq ¼ nd !ð1
rd Þ The number of packets in buffer Nq can be calculated as 1 X rd Nq ¼ kpk ¼ pq ð1
rd Þ k¼n

ð6Þ

ð7Þ

d

Applying Little’s theorem, the mean queuing delay per packet can be obtained as, t¼

Nq ld

ð8Þ

The channel utilization is defined as the fraction of the system capacity utilized at steady state. We have considered a system in which the average data rate is equal to the capacity of each OFDM sub-channel. Hence, the link utilization is the fraction of the average number of channels utilized at steady state which is rd : Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

DISTRIBUTED MAC PROTOCOLS

483

3.4. Computation of statistical packet dropping probability We now derive the delay distribution for packets in buffer and compute the probability that a packet exceeds a predefined delay limit tmax : The delay experienced by a kth packet in buffer follows the Erlang distribution [15] with mean nd md and parameter k: Assuming that there are already k packets in buffer, Wkþ1 ¼ Erfk þ 1; nd md g

ð9Þ

ðnd md Þe
nd md t ðnd md tÞk k! The probability that the queuing delay,Wq exceeds a delay limit tmax is given by 1 X PrfWq > tmax g ¼ pk PrfWkþ1 > tmax g ¼

ð10Þ

ð11Þ

k¼0

Substituting for pk from Equation (3), we obtain 1 Z 1 k
n r nd X rd ðnd md Þe
nd md t ðnd md tÞk d d p0 PrfWq > tmax g ¼ m! k! t k¼0

ð12Þ

Simplifying the above equation, we obtain the statistical dropping probability as PrfWq > tmax g ¼ pq e
nd mð1
rd Þtmax

ð13Þ

3.5. Results and discussion In this section we present the simulation results of CSMA/CCA strategy and a reservation MAC protocol. We compare the simulation results of CSMA/CCA with analysis results. For the reservation protocol, uplink and downlink signaling slot length of 4 ms is chosen. Figure 6 shows the semilog plot of access delay as a function of number of stations. It can be observed that for low loads such as 40 access stations, the delay in DMAC is 7 ms; while the delay in CMAC is 10 ms: It can be inferred that for such low loads, the probability of a packet getting buffered is very low and since there is no signaling delay in DMAC, the total delay is very low in the order of few microseconds. But for the CMAC case, though the probability of buffering a packet or collision of request packets is low, the minimum signaling delay of 8 ms causes the overall delay to be 10 ms: The delay for DMAC is only due to queuing delay whereas the delay for CMAC is due to 3 components namely, queuing delay, request retransmission delay, and signaling delay. For medium loads such as for 120 stations, the queuing delay is in the order of signaling delay and hence the difference in delay between CMAC and DMAC narrows down. It could be seen that for high loads such as 170 stations, the delay due to DMAC is 80 ms while that of CMAC is 90 ms: In these loads, the effect of queuing delay is dominant than the sum of delays due to signaling and collisions. Hence the delay of CMAC is nearly equal to the delay due to DMAC. It should be emphasized that, even if the performance of DMAC is equal to that CMAC in high loads, the performance of DMAC is obtained with very less signaling compared to CMAC. It is also seen that the analysis plot of DMAC matches closely with the simulation plot, thus validating our Markovian model. The Figure 7 shows the statistical dropping probability as a function of number of stations. Assuming the upper limit on statistical dropping probability to be 1% and the maximum Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

484

S. SUNDARESAN ET AL.

10

3

Centralized Distributed-Simulation Distributed-Analysis

Mean access delay (in ms)

10

10

10

10

10

10

2

1

0

-1

-2

-3

0

20

40

60

80 100 Number of stations

120

140

160

180

Figure 6. Access delay plot for CMAC and DMAC with nd ¼ 8 channels.

tolerable limit as 100 ms, it is observed from the graphs of Figure 7 that CMAC supports 100 stations, while DMAC supports 110 stations without violating statistical QoS. Figure 8 shows the channel utilization achieved subject to statistical QoS constraints. It could be seen that the CSMA/CCA scheme has greater channel utilization than the reservation protocol. For a statistical dropping probability of 1%, the utilization achieved by CMAC is 0.49 while that achieved by DMAC is 0.54 as shown by Figure 8. This means that our proposed DMAC achieves 10% increase in throughput compared to CMAC. Hence, our proposed DMAC scheme performs much better than the CMAC scheme in terms of access delay and utilization.

4. PRIORITY ORIENTED SCHEDULING IN DOWNLINK Packet scheduling is important in effective utilization of available bandwidth as well as quality of service (QoS) provisioning [21]. In a shared medium such as powerlines and wireless, scheduling also play an important role in channel allocation that optimizes transmitted power. Packets that arrive from the backbone to the scheduler have various characteristics such as packet size, rate of arrival and burstiness. These properties determine the different levels of Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

485

DISTRIBUTED MAC PROTOCOLS

0

10

Centralized Distributed-Simulation Distributed-Analysis -1

Statistical Dropping Probability

10

-2

10

-3

10

-4

10

-5

10

-6

10

0

20

40

60

80 100 Number of stations

120

140

160

180

Figure 7. Statistical dropping probability for CMAC and DMAC with nd ¼ 8 channels for a threshold of t ¼ 100 ms:

network requirements. The ability of the scheduler to satisfy these requirements is called QoS provisioning. Some of these requirements are maximum tolerable delay, variable bit-rate support and optimum power. Hence, we implement a novel scheduling scheme, in which buffered packets are serviced in the order of their priority and the channel allocation is done using the MaxFlow algorithm. We describe the downlink traffic model and the adaptive modulation of downlink channels in this section. We then describe the proposed scheduling policy and perform extensive simulations. The performance of the system in terms of buffering delay, statistical dropping probability, transmitted power and the number of stations is then analysed.

4.1. Downlink traffic model We classify the downlink traffic into three broad classes. They are 1. Class A traffic: Low file size/constant bit-rate applications E.g.: SMTP(Email), TELNET packets and online interactive sessions. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

486

S. SUNDARESAN ET AL.

0.9 Centralized Distributed 0.8

0.7

Utilization

0.6

0.5

0.4

0.3

0.2

0.1 -4 10

-3

10

-2

10 Statistical Dropping Probability

-1

10

Figure 8. Channel utilization for CMAC and DMAC subject to statistical dropping probability with nd ¼ 8 channels.

2. Class B traffic: High file size/constant bit-rate applications E.g.: Web downloads, file transfers and JPEG images. 3. Class C traffic: Variable bit-rate applications E.g.: Real-time video. The total packet arrival rate in downlink is same as uplink, as every downlink packet is a response to uplink request. Hence the arrival rate of packets per access station in downlink is 0:2 pkts=s: The arrival process is poisson distributed and the packet sizes have an exponential distribution. The statistical properties are summarized in Table II. We denote class A packets as (A,64) and class B packets as (B,64). The class C packets are denoted as (C,64), (C,256) and (C,1024) based on the bit-rate required by the incoming class C packet. 4.2. Adaptive modulation of downlink channels OFDM offers the flexibility of adaptive modulation for time dispersive medium like powerlines and wireless. In such medium, there exists some frequencies (‘good carriers’) with better signalto-noise ratio (SNR) leading to a lower bit error rate (BER) and certain frequencies (‘bad carriers’) that have a very low SNR leading to a higher BER. In a system with adaptive modulation, the carriers are modulated based on their SNR. We recall that a total of 18 Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

487

DISTRIBUTED MAC PROTOCOLS

Table II. Downlink traffic parameters. Class of traffic Class A Class B Class C

Arrival rate

Holding time

Priority

0:6ld 0:3ld 0:1ld

187:5 ms 1:0 s 1:0 s; 10 ms 100 ms

Least Medium Highest

PRIORITY SCHEDULER

Bit rate (s) (in kbps) required 64 64 64 2564 1024

COMMUNICATION LINK

MAXFLOW SCHEME

Figure 9. Priority scheduler with MAXFLOW.

channels are allocated for downlink traffic. In this study, we assume that out of this 18 subchannels, 12 sub-channels are assumed to have moderate SNR and are modulated using QPSK, yielding a data rate of 64 kbps on each sub-channel. Out of the remaining six sub-channels, four sub-channels are assumed to have a better SNR and are modulated using 16-QAM, yielding a data rate of 128 kbps: We assume that the remaining two sub-channels have the best SNR. The adaptive OFDM system modulates these two sub-channels using 256-QAM, yielding a bit-rate of 256 kbps: We assume in this system, that all subscribers see the same signal strength in any channel. 4.3. Priority oriented preemptive scheduling We implement a novel scheduling policy for downlink traffic in a PLC access network in this section. The scheduling policy has two main components namely the priority scheduler and the MaxFlow block as shown in Figure 9. The functions of the two blocks are explained further. 4.3.1. Priority scheduler. 1. The packets are buffered according to their class in respective buffers. 2. The priority scheduler serves the class C buffer first and serves class B buffer only when class C buffer is empty. (A,64) are served on best effort i.e. class A packets are given whatever bandwidth is available that is not used by (B,64) and ðC; * Þ} packets. 3. If the priority scheduler is serving a ðC; * Þ; it checks if it could satisfy the rate required by the ðC; * Þ by preempting some or all of (A,64) or (B,64) packets that are currently undergoing service. 4. If the condition in step 3 is satisfied, then the scheduler preempts the required number of (A,64) and (B,64) packets that are currently in service and queues them back into their }

ðC; * Þ denotes either of (C,64) or (C,256) or (C,1024).

Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

488

S. SUNDARESAN ET AL.

respective buffers by placing them at the head of the buffers. The ðC; * Þ packet is then transmitted without any queuing delay. 5. If the condition in step 3 is not satisfied, the buffered ðC; * Þ packet is delayed till the departure of sufficient (C,64) or (C,256) or (C,1024) packets that are currently in service. After the departure, the condition in step 3 is again verified and the ðC; * Þ packet is transmitted. 6. When the scheduler serves the class B buffer or class A buffer, the transmit epoch of the (A,64) or (B,64) packet is scheduled at the instant when the first departure occurs freeing one or more channels. 4.3.2. MaxFlow algorithm. The channel allocation of the scheduled packets is done using MaxFlow algorithm. Given a set of packets, the MaxFlow block allocates channels such that the instantaneous link utilization is maximized. The MaxFlow equivalent graph for channel allocation is shown in Figure 10. Nodes u1 ; u2 ; . . . represent the users (AUs) to whom the packets are destined and c1 ; c2 ; . . . represent the channels over which the packets are transmitted. ‘s’ is the source node and ‘t’ is the sink node. It is noted that the network without the source and sink forms a bi-bipartite graph i.e. the vertex set can be partitioned into V ¼ U [ C; where U and C represent the set of users and set of channels respectively. For any ui ; uk 2 U ; and for any ck ; cl 2 C; ðui ; uk Þ 2= E and ðck ; cl Þ 2= E; where E is the set of edges in the network. This means that the set of vertexes U and C are disjoint. The capacities of the edges are as follows * * *

The edges of the form ðs; uk Þ has capacity, Cðs; uk Þ ¼ 1 8k The edges of the form ðui ; cj Þ has capacity, Cðui ; cj Þ ¼ 1 8i; j The edges of the form ðcl ; tÞ has capacity equal to the maximum possible data rate on channel cl :

u1

c1 C=64 c2

F=64

s

C=64

u2

F=256 C=128

t

c14 F=64

C=256

uk

c18

Figure 10. A Bipartite graph illustrating the scheduling scheme. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

489

DISTRIBUTED MAC PROTOCOLS

The flow incoming to an user vertex in the graph is satisfied by choosing appropriate paths between the source ‘s’ and sink ‘t’, such that the flow conservation constraint and capacity constraint [24] is not violated. Once a path is chosen, to partially or completely satisfy a flow incoming to an user vertex, the edges between the channel vertex on the path to other user vertexes not lying on the path are deleted. This is to ensure that not more than one packet is simultaneously transmitted over a channel. Similarly, when the flow incoming to an user vertex is completely satisfied, the edges between the user vertex and the channel vertexes which are not assigned to the user are deleted. The channels with moderate gain which support a maximum of 64 kbps are assigned channel indices between 1 and 12. Similarly those channels with better gain which support a maximum of 128 kbps are assigned channel indices between 13 and 16. The channels with best gain supporting a maximum of 256 kbps are assigned channel indices of 17 and 18. The bipartite graph of Figure 10 represents the snapshot when the MaxFlow block has to allocate channels for (A,64),(B,64) and (C,256) packets. The free channels that are available are c1 ; c2 ; c14 ; c18 : The following steps describe the functions of the MaxFlow block along with illustrations. 1. When a (C,256) or (C,1024) arrives, the MaxFlow block chooses free channels of highest available index. With regard to Figure 10, the flow in the edge ðc; u2 Þ is completely satisfied by choosing the path fs; u1 ; c18 ; tg: As the flow incoming to u2 is completely satisfied, the edges between u2 and the channel vertexes ðc1 ; c2 ; c14 Þ as well as the edges between c18 and the user vertexes ðu1 ; u2 ; uk Þ are deleted. 2. When a (A,64) or (B,64) or (C,64) arrives, the MaxFlow block chooses the free channel of least available index. In Figure 10, the flow in the edge ðc; u1 Þ is completely satisfied by choosing the path fs; u1 ; c1 ; tg: Similarly, for the flow in the edge ðc; uk Þ; the next free

u1

c1

F=64

s

F=

u2

64

c2 F=64 F=256

F=

C=128 F=0

uk

t

c14 C=256 F=256

256

F=64

C=64 F=64 C=64 F=64

c18

Figure 11. Residual graph network after channel allocation. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

490

S. SUNDARESAN ET AL.

channel of least available index is chosen, which is the path fs; uk ; c2 ; tg: As mentioned in earlier, the invalid edges after the channel allocation are deleted. Figure 11 shows the residual graph after channel allocation. From the MaxFlow theory [24], we recall the following attributes for the residual graph network: 1. Net flow F ; is defined as the sum of flows terminating at the sink. The net flow for the network in Figure 11 is 384. 2. A cut S; is defined as the set of vertexes that includes the source node ‘s’ but does not include the sink node ‘t’. There can be more than one cut in a network. The capacity of a cut, CðSÞ is defined as the sum of capacities of the edges that originate at a vertex inside the cut and terminate at a vertex outside the cut. For example, the cut fs; u1 ; u2 ; uk g in Figure 11 has infinite capacity. It can be recalled that for packets which require a bit-rate of 64 kbps; the free channel of least available index is chosen. If all the 12 sub-channels of 64 kbps are busy, and if the arriving packet’s priority is such that it cannot preempt any packet currently in service, the channels that support higher bit-rates (which are c13 through c18 ) are allocated to the arriving packet. In such a case, these channels are modulated using QPSK as the packets require 64 kbps only. Since these channels are of higher gain, the required energy per symbol time is less for QPSK modulation than the case when the channels are modulated to their maximum capacity. If ‘Ek ’ is the energy required to transmit ‘k’ bits per symbol time over a channel, the energy required to transmit ‘k1 ’ bits per symbol time over the same channel and for the same BER, is given by the relation, E k1 2k1
1 ¼ k Ek 2
1 Hence, for a fixed BER, if E is the energy required to transmit 4 bits over an OFDM symbol time (using 16-QAM) in the channels c13 through c16 ; then the energy required to transmit 2 bits (using QPSK) over the same channels is E=5: Similarly, if E is the energy required to transmit 8 bits over an OFDM symbol time (using 256-QAM) in the channels c17 through c18 ; the energy required to transmit 2 bits (using QPSK) over the same channels is E=21: This reduction in energy while transmitting (A,64) or (B,64) or (C,64) packets over the higher gain channels results in power savings. Hence our scheduling discipline also saves transmitted power. When a (C,256) or (C,1024) arrives, the scheduler chooses free channels of highest available index. If the free channels exist such that the rate requirement is satisfied, the packet is transmitted with energy E per OFDM symbol time. If the rate requirement is not satisfied by the existing free channels or by the preemption of Class A and Class B packets, the scheduler checks if any (C,64) packets are being transmitted in better/best channels. If there exists such (C,64) packets, these packets are relocated to channels with index 1 to 12. Relocation is done either by replacing the better/best channel by a free moderate gain channel or by preempting a Class A or a Class B packet being transmitted over an moderate gain channel. If there are no such (C,64) packets, (C,256) or (C,1024) are queued till sufficient packets complete service, thereby freeing channels. 4.3.3. Results and discussion. We describe the simulation results of the proposed scheduling scheme in this section. The mean delay for class A, class B and class C packets are shown in Figure 12. It could be seen that the average delay of class C type packet is zero. This is the Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

491

DISTRIBUTED MAC PROTOCOLS

5

10

Class A Class B Class C

4

10

3

10

Mean delay (in ms)

2

10

1

10

0

10

-1

10

-2

10

-3

10

60

80

100

120 Number of stations

140

160

180

Figure 12. Mean packet delays versus number of stations with 18 downlink channels and adaptive modulation.

advantage offered by adaptive modulation. It is also observed that class A has a mean delay of 1 s and class B packets have a mean delay of 80 ms for high loads such as 170 stations. The maximum tolerable delay for class A packets is 5 s and that of class B packets is 1 s [20]. It can be seen that 180 stations is the saturation point when the Class A delay increases steeply because of it’s lowest priority. The Figure 13 shows the probability that class A, class B, and class C packets cross their threshold in the system. It can be seen that, only class A packets have their statistical limit to violate 1%. Hence, from the results for statistical probability, it can be inferred that the system supports 160 access stations without violating the offered statistical QoS. As class A packets are delay insensitive, the network can choose not to provide the statistical QoS requirements of class A packets. Hence, the system can support around 170 stations. It can be concluded from the results that the priority scheduling with adaptive modulation satisfies the delay requirements of all types of packets and also achieves higher throughput. Figure 14 compares the transmitted power for the AMOD system and a system in which all channels are modulated using QPSK only. The power calculated here is normalized and do not reflect actual values. It can be noted that for a load as high as 170 stations, the transmit power with adaptive modulation is 26% less than that without adaptive modulation. Therefore, adaptive modulation with priority scheduling policy has the following advantages. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

492

S. SUNDARESAN ET AL.

0

10

Class A Class B -1

Statistical dropping probability

10

-2

10

-3

10

-4

10

-5

10

120

130

140

150 Number of stations

160

170

180

Figure 13. Statistical dropping probabilities versus number of stations with 18 downlink channels and adaptive modulation.

1. 2. 3. 4.

Statistical QoS requirements of all types of packets are satisfied. The number of access stations that can be supported is as high as 170. Results in about 26% power savings. Real time applications experience zero delay.

5. CONCLUSIONS A distributed media access control protocols for uplink and priority scheduling policies for downlink traffic in a OFDM based PLC network has been studied. We have then described OFDM modulation method based on which the MAC and scheduling protocols were studied. We studied a CSMA technique with centralized collision resolution for uplink data traffic. The distributed approach resulted in very less access delay compared to reservation protocol during low and medium loads. But in high loads, the delay performance of our DMAC and the CMAC were same. The DMAC achieved 10% improvement in throughput. The DMAC was modelled analytically using Markov chains. The correctness of the simulation results were verified Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

493

DISTRIBUTED MAC PROTOCOLS

2

10

Average power (in units)

NOAMOD AMOD

1

10

0

10

20

40

60

80

100 120 Number of stations

140

160

180

Figure 14. Normalized power versus number of stations with and without adaptive modulation.

through comparisons with analysis results. Priority scheduling for downlink was studied. The traffic in downlink was classified into three classes. The variable bit-rate packets were given the highest priority. Priority scheduling combined with a graph-theoretical approach called MaxFlow was studied for the system with adaptive modulation. The MaxFlow approach is used for efficient channel allocation that will maximize instantaneous link utilization. In this system, the delay of variable bit-rate traffic is obtained as zero and class A and class B packets experience a very low delay. The system is capable of supporting 160 access stations without violating statistical constraints of class A and class B packets. The adaptive OFDM system also saves transmitted power. This work can be extended by considering pooled resources for voice and data. REFERENCES 1. Phillips H. Development of a statistical model for powerline communication channels. Proceedings of ISPLC, Limerick 2000; 153–160. 2. Prasad TV, Srikanth S, Krishnan CN, Ramakrishna PV. Wideband characterization of low voltage outdoor powerline communication channels in India. Proceedings of ISPLC, Sweden 2001; 359–364. 3. Chuang J, Sollenberger N. Beyond 3G: Wideband wireless data access based on OFDM and dynamic packet assignment. IEEE Communications Magazine, July 2000; 78–87. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

494

S. SUNDARESAN ET AL.

4. Hransica H, Haidine A. Modelling MAC layer for powerline communication networks. SPIE Symposium on Voice, Video and Data Communication, 2000. 5. Hanzo L, Webb W, Keller T. Single- and Multi-carrier Quadrature Amplitude Modulation. Wiley: New York, 2001. 6. Prasad R, Van Nee R. OFDM for Wireless Multimedia Communications. Artech House Publishers: Norwood, MA, 2000. 7. Coffey T, Griffin J, Moore B. A medium access control protocol for a powerline local area network. Proceedings of ISPLC, Tokyo 1998; 179–193. 8. Tsuzuki S, Yamadi Y. Utilization and delay performance analysis of carrier sense CDMA protocol. Proceedings of ISPLC, Limerick 2000; 80–89. 9. Stancheva M, Begain K, Hransica H, Lehnert R. Suitable MAC protocols for an OFDM based PLC network. Proceedings of ISPLC, Limerick 2000; 241–248. 10. Langguth T, Steffen R, Zeller M, Steckenbiller H. Performance study of access control in powerline communication. Proceedings of ISPLC, Limerick 2000; 97–104. 11. Mangold S, Matheus A, Beine M. Powerline Communications: Adopting Protocols of Wireless Access Networks. Africom, Cape Town, 2001. 12. Jeong D, Jeon W. CDMA/TDD system for wireless multimedia services with traffic unbalance between uplink and Downlink. IEEE-JSAC, May 1999; 939–946. 13. Chuang J, Sollenberger N. Spectrum resource allocation for wireless packet access with application to advanced cellular internet service. IEEE-JSAC, August 1998; 820–829. 14. Bertsekas D, Gallager R. Data Networks (2nd edn). Prentice Hall of India Private Ltd, 2000. 15. Hillier F, Lieberman G. Introduction to Operations Research. McGraw Hill: New York, 1995. 16. Zimmerman M, Dostert K. An analysis of broadband noise scenario in powerline networks. Proceedings of ISPLC, Limerick 2000; 131–160. 17. Rapport TS. Wireless Communications, Prentice Hall: Englewood Cliffs, NJ, 1999. 18. Ross S. Introduction to Probability Models (6th edn). Harcourt Asia Pte Ltd: Singapore, 2000. 19. Choi H, Limb J. A Behavioral Model of Web Traffic. Intl Conf on Network Performance, 1994. 20. Gurbuz O. Dynamic Resource scheduling schemes for W-CDMA systems. IEEE Communications Magazine, October 2000; 80–84. 21. Keshav S. An Engineering Approach to Computer Networking (2nd edn). PHI Ltd: Singapore, 2001. 22. Krunz M. Bandwidth allocation strategies for transporting variable-bit-rate video traffic. IEEE Communications Magazine, January 1999; 40–46. 23. Sen P, Maglaris B, Rikli N, Anastassiou D. Models for packet switching of variable-bit-rate video sources. IEEEJSAC, June 1989; 865–868. 24. Cormen TH, Leiserson CE, Rivest RL. Introduction to Algorithms. MIT Press: Cambridge, MA, 1995. 25. www.intellon.com 26. www.ascom.de 27. www.inari.com 28. www.itrancomm.com

AUTHORS’ BIOGRAPHIES

S. Sundaresan was born in India. He received the BE Degree in Electronics and Communication from Madurai Kamaraj University, India. He has submitted his thesis for MS in Anna University and is awaiting his defense. He was a Part-time research assistant in AU-KBC research center and was part of research in powerline communications. He is currently Member Technical Staff in HCL Technologies, Chennai, India. His field of specialisation is MAC and scheduling protocols for powerline and wireless networks. His research interests include networking, wireless systems and OFDM.

S. Anand was born in India. He received the BE Degree in Electronics and Communication from Anna University, Madras, India in 1997 and ME Indian Institute of Science, Bangalore, India 1999. He is currently pursuing his PhD in Indian Institute of Science, Bangalore, India. His research interests include medium access protocols, channel allocation and scheduling policies for CDMA/OFDM based networks. He has also published many journals in the area of CDMA based wireless communications. Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495

DISTRIBUTED MAC PROTOCOLS

495

S. Srikanth was born in India. He received the BE Degree in Electronics and Communication from Anna University, Madras, India in 1990 and MA Sc from University of Victoria, Canada in 1993. He completed his PhD from University of Victoria, Canada in 1997.He was with Harris Corporation, Canada as DSP engineer till 1999. He was a Visiting Professor to College of Engineering, Anna University till 2000. He is currently holding the position of Member Research Staff in AU-KBC research center. His field of specialization and interests include communications, space-time coding and multi-carrier techniques for mobile communication systems. C.N. Krishnan was born in India, in 1947. He received the BTech degree from The Indian Institute of Technology (IIT), Madras, Indian in 1969 and MTech degree in electrical engineering from the IIT, Kanpur, India in 1971. He completed his PhD in the area of microwave semiconductor devices in IIT, Kanpur in 1977. He was a Senior Research Assistant in IIT, Kanpur till 1977. Since 1977 he has been with Madras Institute of Technology, Anna University, India, where he is currently a Professor of Electronics. He is also the Director of AU-KBC research center in which the research on powerline communications is being done. His field of specialization and interests include communications, signal processing and networking. He has undertaken numerous funded research projects in the areas such as CDMA, GPS, and Multimedia, etc.

Copyright # 2003 John Wiley & Sons, Ltd.

Int. J. Commun. Syst. 2003; 16:475–495