A Wireless MAC Protocol comparison.

Transcription

1 Doc: IEEE P S1 IEEE Wireless Access Method and Physical Layer Specifications Title: Presented by: A Wireless MAC Protocol comparison. Wim Diepstraten NCR SEUtrecht NCRAT&T Network Prodct Grop Niewco.:r: The Netherlands (V) (Fax) Wim. Diepstraten@ U trech t. ncr.com ( ) Abstract: In this paper a several protocol comparison aspects are addressed. An analyses is provided that shows the obvios advantage of a distribted access protocol. The W A VELAN CSMNCA protocol is explained and its performance is compared against the 4WAY LBT protocol as proposed by Ken Biba, and against an extension of the crrent Wavelan protocol with a MAC level recovery. Several performance analyses methodologies are proposed to bild a common grond for protocol comparison. The presented reslts are preliminary and are based on the simlation approach described in IEEE P Conclsion: It was conclded that the interference robstness and medim sharing behavior are the most important characteristics of a wireless protocol. Based on this a global analyses shows that a distribted "coordination fnction" is best sited to achieve atomatic sharing of the available medim bandwidth, withot additional coordination overhead. To improve the robstness of the protocol a MAC level recovery mechanism is very effective. In addition it is shown that CSMNCA as sed by W A VELAN does have a very high throghpt efficiency paired with a low transfer delay, which is similar than the perfonnance of 82.3 CSMACD networks. Wavelan CSMACA, CSMACA + Ack and the modified Ken Biba 4way LBT protocol are Printed: MAY 8, 1992 Pge 1 By: Wim Diepstraten

2 Doc: IEEE P compared on several aspects. A PeerToPeer Bffered Load test methodology is proposed and allows protocols to be compared to several relevant characteristics. The relevant comparison methodologies are explained and illstrated with several simlation otpt charts. A ClientServer test configration is sggested to evalate the net performance for an application. Frther work will be needed to compare the different protocols, and this will be sbject for ftre contribtivns. Introdction: When analyzing MAC protocols, the qestion is what kind of characteristics shold be looked at an, what wold be the sitable method to do it. Clearly one important aspect will be the throghpt and response times for the individal stations, and the capacity of the whole system. In a wireless environment in particlar, the robstness for interference, and the ability to share the medim are very important characteristics which need to be analyzed when comparing different MAC protocols. This docment addresses the important characteristics, and proposes a methodology to analyze them. Only the asynchronos services are evalated. As an example_ three possible protocol implementations are compared against each other. The following protocols are sed as an example: CSMACA protocol as sed by W A VELAN. CSMACA + Ack protocol extension. 4WAY LBT protocol (a modified Ken Biba proposal) The intention of this paper is not to arrive at a conclsion abot the preferred access method, becase to date not enogh simlations are done to flly compare all characteristics of the protocols. Protocol Characteristics: The main characteristics of a protocol will depend on the type of services that the MAC layer needs to spply to a particlar application. We can distinct a few different categories of applications that se two different types of services as follows: Connectionless service. This is the type of service which most of today's wired LAN's provide, and on which the majority of Network Operating Systems and applications are based on. It is ideal for "Brsty" traffic. It is characterized by a very low response time, and is generally very tolerable for large deviations in transfer delay. Printed: MAY Page 2 By: Wim Diepstraten

3 Doc: IEEE P Connection oriented service. This is a type of service that can garantee a fixed bandwidth to an application, and is characterized by a relative long setp time for the connection, after which the data transfers are sally providing a more or less constant bit stream with relative low deviation in transfer delay. This type of services are sed for applications which need real time data like voice or video, and for time bonded transfers of data in for instance an indstrial environment. This docment will primarily deal with the connectionless type of service, most commonly tilized by the majority of today's Network Operating Systems. The following is a list of characteristics that need attention: Throghpt: Total system throghpt Individal station throghpt as fnction of Network load. How is throghpt shared with other Networks sing the same channel. Effect of interference: From other Networks. From other sorces. Transfer Delay at varios load conditions. Max nmber of stations. Fairness. BehaviorStability nder havy load conditions. BehaviorStability nder varying fade conditions. Effect of "Hidden" tenninals. Effect on higher layer protocols. As conclded in IEEE P simlations will become very important to analyze the perfonnance aspects in a complex environment like Radio Lan's. The relevant Radio medim characteristics need to be modeled sch that their major impact on the performance and operational robstness are factored in. This is done in the described simlation approach. The reslts of the simlations need to be analyzed in sch a way that the major protocol characteristics as listed above are covered. Printed: MAY Page 3 By: Wim Diepstraten

4 Doc: IEEE P S1 Protocols considered: In general two access protocol implementation categories can be distingished, by the type of "Coordination Fnction" which is applied. The two "Coordination Fnction" types are: Centralized Distribted When the "coordination fnction" is centralized in natre then there is some mechanism that assres, at least in one network segment, that only one station is allowed to access the medim at anyone time. The "coordination fnction" can itself be located in a fixed station, like in a polling environment, or it can be traveling throgh all stations within a network segment (BSA) like in a token passing system. When the "coordination fnction" is distribted in natre then there is a procedre defined by which individal stations are trying to access the medim resolve the contention on the medim. This applies to _ the CSMA type of random access protocols. A wireless medim is very different from a wired medim, and most obvios is the channel separation difference and availability of bandwidth. In a wired environment the separation between different physical networks is infmite, becase they rn on separate cables. Increasing the system capacity can simply be achieved by increasing the system bandwidth by rnning extra cables. In a wireless environment however spectrm resorces are scarce. The conseqence is that at worst mltiple networks will need to share the same band. The isolation between the different network segments (BSA's) will be dependent on the nmber of channels available. In a mlti freqency channel system the isolation can be relatively high and will depend on the spatial rese distance that can be achieved with the given nmber of channels available. When mltiple overlapping channels are sed then the channel separation will depend on the transmitter and receiver design (PRY). The sprios emissions of the transmitter in a neighbor channel, and the non linear distortion generated in the receiver by a strong nearby transmitter are casing this interference. When only one freqency band is available then an other method for channelization can still be applied by sing different (orthogonal) codes in a direct seqence spread spectrm system. The achievable isolation will depend on the code length sed, and will be the tradeoff between raw bitrate and isolation. In practice for bitrates higher than 1 Mbit the achievable isolation will be very poor compared to the dynamic range of the radio signal, Printed: MAY 8,1992 Page 4 By: Wim Diepstraten

5 Doc: IEEE P and accrate power control mechanisms are needed to make it sefl. In addition not many orthogonal codes will be available. Bottom line, the cochannel interference in these systems can still be significant, and needs carefl consideration for the protocol design. The lowest isolation will be the system which has only one channel available withot any frther separation provisions. Here the same medim mst be shared by mltiple networks nless sfficient spatial isolation can be achieved. In all three cases the protocol has to deal with a certain cochannel interference level. Since the protocol mst be able to rn on different PHY's, the worst case sitation needs to be considered, being the single channel environment. Please note that one of the reqirements listed in the Draft reqirement docment was the ability to spport a single channel environment. Therefore the coverage area of a wireless LAN segment (BSA) will be interference limited rather than noise limited. The interference tolerance of a protocol specifically the tolerance for cochannel interference will be a very important factor in the MAC protocol design. This will translate into two different aspects: The access mechanism: The "coordination fnction" shold be robst for interference. The transfer sccess: The interference sitation at the intended receiver station will determine the sccess of the data transfer. Since it will be very hard to predict this from the transmit station or any other "coordination fnction" location, especially in the case of other (non cochannel) interference sorces, the sccess of a transfer can not be garanteed. Hence the protocol shold be tolerant for lost packets. The latter can best be achieved by a recovery mechanism on the MAC. to become independent from the recovery mechanisms sed at higher protocol levels. This is becase those recovery procedres are not designed for an environment where interference is dominating the packet error probability, so they are less efficient. A positive acknowledge mechanism, which will indicate to the transmitting station that the packet has been sccessflly received. can be sed to detect that packets are lost. For the access mechanism the basic choice will be between the centralized and distribted "coordination fnction". In my mind the determining factor affecting this choice is primarily the ability and efficiency to share the same medim with other networks. For "brsty" traffic a centralized "coordination Printed: MAY 8,1992 Page 5 By: Wim Diepstralen

6 Doc: IEEE P S1 fnction" will not be able to efficiently coordinate the medim access between mltiple BSA's, especially when those BSA's are not part of the same ESA. It wold involve coordination via the distribtion system if any exist more or less on a per packet basis. This type of coordination cold perhaps be dealt with for connection oriented traffic, bt still a big problem remain between mltiple ESA' s. In any event this type of "coordination fnction" will be pretty complex and relatively inefficient (for brsty traffic). This is where a distribted "coordination fnction" can be very efficient becase it provides for atomatic sharing of the medim withot any added complexity (for the coordination). A frther advantage is that at a low average network load which is typical for most LANs today, the transfer delay is very low and provides for a "snappy" response. An other factor which needs to be considered is the amont of bandwidth needed for the "coordination fnction" on the Wireless medim and on the distribtion system (backbone network). To give an example: a token passing scheme wold continosly se scarce bandwidth even when no load is applied to the system. This will seriosly effect the sharing of the channel with neighbor BSA's. Coordination Fnction type Conclsion: The main characteristic for a Wireless MAC protocol is its robstness for interference, and related to that the ability to share a single medim with other networks efficiently. A global analyses shows that for "Brsty" data traffic typical for the connectionless services sed by most Network Operating systems today, a distribted "coordination fnction" promises low response times and more or less atomatic sharing of the medim between overlapping networks. This allows for a gracefl degradation in throghpt per individal network. Apart from this access mechanism it will always be necessary to have a method of dealing with lost packets, and a MAC level recovery mechanism will therefore be essential to achieve adeqate robstness. To allow the implementation of a MAC level recovery mechanism a lost packet detection mechanism is needed, which can be done with a relative simple positive Acknowledge facility. Printed: MAY 8,1992 Page 6 By: Wim Diepstraten

7 Doc: IEEE P Possible "distribted" access methods. In wired LAN environments CSMNCD is accepted to be a very effective medim access method. In a radio environment it is however not practical to implement the CD fnction becase of the extreme high dynamic range reqired. The WAVELAN prodct on the market today ses CSMNCA, which is CSMA combined with a collision avoidance strategy. This provides for a medim se efficiency near the CSMACD efficiency. The 4WA Y LBT protocol proposed by Ken Biba is also a distribted type of access protocol. CSMNCA explained: The main access mechanism is Ipersistent CSMA. First of all the medim is sensed for valid Spread Spectrm carrier, and when silence is sensed dring the dration of a slot, the medim will be accessed. When the medim is sensed bsy, then the transmitter will defer ntil the end of the packet. Normally in CSMA a deferring station wold access the medim immediately after the IFS (inter frame space) period. At medim to high network load, there is however a high probability that more stations were deferring on the same packet, and when they both access the medim after the IFS period this will reslt in a collision, assming that at the intended receivers both signals will interfere with each other. This is the sitation which CSMNCA is trying to avoid. A deferring station will at the end of the packet start a random backoff seqence which is designed to significantly redce the probability that the mltiple deferring stations will collide. So when a station senses a bsy network it will se ppersistent CSMA right after the IFS. CSMNCA Performance: Figres 1 compares the W A VELAN CSMACA performance with ALOHA and slotted CSMA which is rn on the same W A VELAN PHY. As shown, the efficiency of the protocol is very good, and reaches approximately 87% of the 2 Mbps raw bitrate for long packets. This is similar to the efficiency of CSMNCD parameterised according to the 82.3 standard. It shold be noted that part of the overhead is de to the training time reqired by the PHY to select the best antenna (for antenna diversity), and achieve proper synchronization. The slotting time relates to the time reqired for proper Spread Spectrm signal detection on a particlar antenna. Given the low medim propagation delay, the slotting time can be mch lower than sed in the 82.3 CSMACD networks today. For Wavelan the slottime is in the order of 1% of the 82.3 standard. This clearly shows that a distribted access method can achieve high efficiency, paired with low transfer delay, and is stable even nder high load conditions. Printed: MAY 8,1992 Page 7 By: Wim DiepstrJten

8 Doc: IEEE P S1 Comparing MAC protocols. The W A VELAN CSMNCA access method is sed frther to compare it against possible alternatives. One clear alternative is to extend the CSMACA access method with a MAC level recovery mechanism based on a packet loss detection by means of an Acknowledgement of sccessfl reception of the MSDU by the intended receiver. The other distribted control protocol nder analyses is the 4WA Y LBT protocol as proposed by Ken Biba in Doc IEEE P The described protocol has been adapted in certain area's, and is rn on the Wavelan PHY, so that a direct comparison between the MAC protocols is possible. The intention of this paper is not to arrive at a conclsion abot the preferred access method, becase to date not enogh simlations are done to flly compare all characteristics of the protocols. The rest of the paper is intended to describe the possible methodology for comparing the protocols on several aspects. 4WAY LBT Protocol modifications: The 4WAY LBT protocol proposal of Ken Biba in Doc IEEE P is changed in some area's as explained below: The RTS, crs, DATA and Ack frame strctres are changed a bit in size to reflect the following: The preamble is effctively longer to reflect the training length reqirements of the W A VELAN PHY. The crs and ACK packets contain the sorce address of the original transmitter that is captred by the intended receiver. The docment was nclear abot the behavior of the receiver when it receives arts packet. The docment sggest that when a station is receiving a valid RTS addressed to that station, then it will send back the CTS packet. This is changed, so that the crs packet is only retrned when the station is enabled to transmit by the "Net allocation vector", to prevent it from interfering with ongoing transfers in a nearby network. Printed: MAY Page 8 By: Wim Diepstraten

9 Doc: IEEE P Simlation environment: The simlation program as described in doc IEEE P was extended to inclde the modified 4WAY LBT protocol from Ken Biba. All the different protocol derivates shown here are controlled by setting different parameters in the simlation program. In addition some changes were made to the reporting facilities, and the reslts are imported into SigmaPlot version 4.1 and processed to provide the graphical otpt as shown in the appendix. This allows for postprocessing of the large amont of simlation reslts, and makes it possible to correlate all kind of sitations with the different parameter sets sed. Simlations performed: The three protocols CSMNCA, CSMNCA + Ack, and 4WAY LBT are being compared against each other with varios type of simlations: Throghpt and Transfer delay verss Bffered Load. with 1% Long packets (Novell length). with a 6% Short packet + 4% Long Packet mix (Novell lengths). Throghpt and transfer delay verss the nmber of stations. Max throghpt as fnction of the PHY preamble length. Medim sharing behavior (at high load). Throghpt and Transfer delay verss distance between two networks. Throghpt and Transfer delay verss distance between two networks in a mlti floor environment. Individal station behavior in varios mlti network environments. "Hidden Station" behavior. Throghpt in a high load Novell environment (Client Server). All simlations are based on the W A VELAN PHY which provides a 2 Mbps raw bitrate data channel. Printed: MAY Page 9 By: Wim Diepstraten

10 Doc: IEEE P S1 All simlations were done with a network of 7 stations scattered across an environment as follows: Scale = 1 meters per dot. In the above configration station is a Server station, which is sed only dring the Client Server tests. The attenation coefficient is set to 3.5 which is typical for most semi open office environments. Dring this test all commnication is PeertoPeer with random destinations per packet. The throghpt measred dring the PeertoPeer test incldes only the sccessfl received packets and incldes the standard MAC overhead of sorce, destination, length and CRC. It does not cont the overhead of the Ack packet in CSMACA and the RTS, crs and Ack packet in the 4way LBT protocol. Test Configrations sed: The basic test configration sed is the PeertoPeer test, in which the stations generate traffic with random destinations. As described in doc IEEE P this reslts in different path attenation for every individal link, and effects of captring and cochannel interference are inclded in the model. To investigate the inflence of the MAC characteristics on higher protocol layers, a CLient Server test will be performed. In particlar the handshaking as sed within a Novell network operating system environment is sed dring these tests. For that reason the sed Short and Long MSDU lengths are based on the packet sizes sed in a Novell environment. To allow comparison between the different test environments, these packet lengths are also sed in the PeertoPeer test environment. Printed: MAY 8,1992 Page 1 By: Wim Diepstraten

11 Doc: IEEE P Bffered Load: In a simlation environment it is not practical to apply a Poisson distribtion fnction to generate traffic for the medim. Also in a real "Brsty" traffic environment this approach is not realistic. In the simlation we have to deal with a (fixed) nmber of stations that want to get a certain amont of data throgh the medim. In real applications the actal traffic sorce is a transaction or a large file which take a nmber of packets to complete a commnication session. In reality a new packet will not be generated ntil the previos packet has been transferred. The MAC protocol needs to resolve the contention between all the stations who have packets to transmit. This is the load offered to the "coordination fnction", and is what I called the "Bffered Load". The bffered load is generated and calclated as follows: Traffic generation model:! =======================! ! xxxxxxx! Random delay Access delay Xmit delay Random delay ++ Access delay xx Xmit delay Used to control the "bffered load" dring the test, It IS generated randomly with a fixed minimm time of in this case 1 msec. This represents the minimm processing time that is needed to generate a new packet by a station. This will be the time reqired for the MAC to gain access to the medim. This is the time reqired to transmit the total packet, at a given raw bitrate. The bffered load per station is = l(a verage Random + Xmit) * Average Packet length For the total system the bffered load will be the bffered load per station times the nmber of active stations. In words the bffered load represents the average amont of data waiting to be transmitted across the medim. It wold be the maximm achievable throghpt when the access delay is zero. It shold be noted however that for the protocols which have a MAC level recovery fnction, the access time incldes the extra time needed for this recovery. The advantage of this approach is that the transfer delay in an overload sitation does still have a meaning and shows the actal average response times achieved at this network load level. It also gives a good overview of the stability of the protocol Ullder high load conditions. Printed: MAY Page 1 1 By: Wim Diepstraten

12 Doc: IEEE P S1 I wold like to propose that this "Bffered Load" definition is sed to prodce the Throghpt and Transfer Delay verss Bffered Load crve. Throghpt and Transfer delay verss Bffered Load crves: Figre 1 compares the performance of CSMNCA with slotted Ipersistent CSMA and ALOHA for a lk data + Novell + MAC overhead packet size of 188 Bytes to set a reference. Figre la is the same with somewhat adapted load axis to show the stability at high overload conditions. Figre 2 and 2a compares the performance of CSMNCA with CSMACA + Ack and the modified 4WA Y LBT protocol of Ken Biba. Figre 3 compares the 3 protocols in a more realistic network environment which is typical for Novell traffic. The load consist of 6% Short Novell(+MAC overhead) packets of 64 Bytes, and 4% of Long Novell Packets based on a data contents of.5 KByte, reslting in 576 Bytes. CSMNCA shows a higher throghpt than both other protocols, bt it shold be noted that the lost packets are not recovered by the MAC itself so this mst be resolved at higher protocol levels. This will effectively reslt in a lower performance when this is taken into accont compared to the CSMNCA + Ack protocol. The reslts show that the 4WAY LBT has a lower perfonnance which is de to the higher MSDU overhead, and the effect of the PHY training overhead which will be shown later. Throghpt and transfer delay verss the nmber of stations. An other relevant parameter in the behavior of an access protocol is the total nmber of stations for which the contention needs to be resolved. This is especially relevant when those stations apply a high load brst to the network. The probability of sccessfl access resoltion in a distribted access protocol will decrease when the nmber of stations accessing the medim at the same time increases. Typically a network is dimensioned sch that the average load is relative low, so that on average a very good response time is achieved. The probability that at a certain point in time n stations are accessing the medim simltaneosly will depend on the application, and will be very low for high vales of n. In other words: the probability that the protocol needs to resolve access contention between 1 stations in a network with 1 stations will generally be very low. Iowever in a single channel wireless environment where all BSA's need to share the same band, the total poplation of stations which "hear" each other is larger then the nmber of stations belonging to the same BSA. This will depend on the propagation characteristics of the Printed: MAY 8,1992 Page 12 By: Wim Diepstraten

13 Doc: IEEE P S1 environment, and the total station density. The relation between throghpt, delay and nmber of stations doing simltaneos access needs to be analyzed. This is done in figre 4 with the diagram showing the total throghpt and the delay as fnction of the nmber of stations. In addition to that, Figre 4a shows the percentage lost packets encontered. In CSMNCA those packets are not recovered, so this neeas to be resolved on higher protocol layers, reslting in a lower average response time as shown in the diagram. This is also the reason why CSMNCA and one of the reasons why the 4WA Y LBT protocol show a lower throghpt performance and a higher transfer delay. Delay distribtion: Crrently the simlation program does not have sfficient provisions to evalate the delay distribtion. It does average the delay on a per station basis, and this is again averaged on a per network basis as shown in the varios figres. This is however one of the characteristics that need to be analyzed when comparing protocols. Frther simlation program development and simlation effort is needed in this area. Dependency on PHY The main parameter which will have inflence on the MAC protocol performance is the preamble length needed to allow the PHY to achieve sfficient qality of service. This is in effect the training time needed in the PHY to obtain sfficient synchronization for a good qality demodlation. This will particlar have a large inflence on High speed systems. When for instance an eqalizer needs to be trained to allow the high modlation rates, then this will need extra preamble time in the MAC. Even when the absolte time to achieve synchronization stays the same when increasing the data rate of a system, will increase the overhead percentage, becase the overhead expressed in bittimes will increase with the same ratio. To analyze the dependency on this a simlation is needed that shows the throghpt as fnction of the PHY related preamble time. This simlation is done at high load sitations, becase there the effect is dominant. This is shown in figre 5. As is obvios from the diagram the 4WAY LBT protocol is highly effected by the preamble length, becase one MSDU transfer translates into 4 packets on the medim, which each need to train the PHY. Printed: MAY 8,1992 Page 13 By: Wim Diepstraten

14 Doc: IEEE P Medim Sharing Behavior In wire1ess the mlti network behavior is essential especially when only one freqency band is available. It is important to share the (single) medim as efficient as possible, also when mltiple ESA's overlap. Related to that is also the fairness of access that individal stations experience. This shold be evalated for both the flly and partly overlapping case. For this prpose a simlation is set p for two networks, in which the distance between both networks can be varied. Becase the simlation program takes into accont every individal attenation path between all the stations of the two networks, the cochannel interference effects are atomatically inclded. The simlation configration ses 2 networks of seven stations each, with variable distance between the networks. In addition an extra attenation can be specified between the networks to simlate that both networks are on a different floor, or in different area's separated by a thick wall. For each of the three protocols 2 sets of simlations are rn. With a additional separation of and 2 db. With varying the distance between the networks from 5 to 65 meters in steps of 5 meters. A 6% Short and 4% Long MSDU mix is sed throghot the tests. Test dration is 5 seconds per point. This way all possible sitations can be evalated. It does contain both flly and partly overlapping sitations, and contains "Hidden Station" sitations. For instance the test configration with 2 db network separation offset in particlar does contain a lot of "Hidden Station" sitations. The reslts need to be looked at in several ways to evalated the different aspects: Throghpt and delay verss network separation distance. Figres 6 This gives a global overview of the total capacity and how effectively it is shared. Also shown on the same page is the nmber of "Data" packets lost per individal station over the total test period of 5 seconds. However if part of the stations have a losy performance becase the access scheme is nfair, this wold hardly show p on the total throghpt perfonnance, becase then the bandwidth wold be sed p by other stations. Individal station performance. The total network throghpt performance does not sfficiently show all the effects of the network overlap, and the fairness between the stations. For this prpose the nmber of Printed: MAY Page 14 By: Wim Diepstraten

15 Doc: IEEE P S1 sccessflly transferred packets per individal station is given in Figre 7. They show how fair the bandwidth is divided over the different stations in the flly overlapping case. In the partially overlapping case there will be a large deviation between stations, becase some of the stations need to share with stations of both networks while other stations do not see the other network or see only part of the stations of the other network. For the CSMNCA protocols the case of the errors can be analyzed, by looking at the nmber of collisions and nmber of overjammed packets per station. For the 4W A Y LBT Protocol two other vales are of interest which give eqivalent information. These are the nmber of lost RTS packets, which is eqivalent to the collision cont, and the nmber of times no crs packets cold be retrned becase the destination station was not enabled to transmit. This is shown in Figre 8 and 8a. The detailed information is needed to get a good feeling for the most dominant factors inflencing the individal performance. They can also be sed to optimize the different parameters of a given protocol. Reslt evalation: As is shown in the total network throghpt graphs, there is no significant difference in the sharing behavior between the 3 evalated protocols. A difference wold be expected however for the 4WA Y LBT protocol, becase this protocol effectively prevents any parallel transmissions in an area arond both the transmitter and the receiver, while the CSMNCA protocols do this only arond the transmitter. The difference can possibly be explained by the setting of the CRS level and minimm reliable receive level. When looking at the individal station throghpt in figre 7, then it shows that there is less variation between the stations in the 4WAY LBT protocol than the CSMNCA protocols. Frther analyses of the reslts are still needed for a good comparison. A preliminary conclsion is that the extra overhead involved in the 4WAY LBT protocol to prevent the loss of a "Data" packet is less efficient then CSMNCA + Ack. So accepting loss of bandwidth of a long "Data" packet is and retransmit the packet is more bandwidth efficient than the continos overhead for every packet. Printed: MAY 8,1992 Page IS By: Wim Diepstraten

16 Doc: IEEE P Throghpt in a Novell environment: A separate class of simlations will be needed to evalate the effect of the MAC protocol on higher level protocol layers, and to investigate the performance nder several traffic load conditions. One of sch traffic models is inclded in the simlation model and represents the Novell Perform3 test. This test pts a high load on the network, and models the Netware transport protocol NCP (Network Core Protocol). This test is very representative for a ClientServer environment in which all traffic is between a station and a server station, with Reqest!Response handshaking per packet. The total throghpt performance is shown in figres 9 and 9a, which in this case show the net application data, so withot the Novell handshaking overhead and MAC overhead. In sch a configration the effect is that all the traffic in a network rns via the Server, which means that the server typically generates 5% of all the packets. The conseqence is that the throghpt relations and the error probability can be mch different from the figres reslting from the PeertoPeer test configration. This is becase the bffered load and the average nmber of stations who are simltaneosly accessing the medim is less than 5% than in the PeertoPeer configration. This is becase on average more than half the stations are waiting for a response of the Server station before they can generate the next MSDU. This difference is also demonstrated in doc IEEE P , which shows the throghpt and lost package probability as fnction of the nmber of stations for both a PeertoPeer and ClientServer configration. A comparison of the individal station performance is shown in figres 1 and loa, which shows only the nmber of packets received from the server. An other aspect that is the effect of a lost packet on the higher layer protocol. This is however only relevant when this higher layer protocol needs to recover from this. In the case where the lost package detection and recovery is done in the MAC, like in the CSMNCA+Ack and 4WAY LBT protocol, this wold be mch less of an isse, and the reslt wold only be a longer response time for that station, so a bit lower throghpt. The effect of this is very visible in the individal performance of the CSMNCA protocol, which need recovery from the higher protocol layers. The effect is a pretty large deviation in individal performance, becase of the long timeot period sed in the higher layers to detect the occrrence of a lost packet. The protocols with MAC recovery do however increase the "Bffered Load" on the medim when many lost packets need to be recovered. This is an effect which is not visible in a Peerto Peer test configration, bt becomes more visible in a ClientServer test. An other aspect that will impact the reslt is the physical location of the Server, especially in a mlti network test. It will make a diference when the Server is located in the middle or near the edge of a network. Prinled: MAY 8,1992 Page 16 By: Wim Diepstraten

17 Doc: IEEE P Bottom line is that we have to look at the individal station performance rather than the total network throghpt performance. This is becase a poor perfonning station will have a low impact on the total throghpt, becase other stations will se the bandwidth that becomes available when a poor perfonning station is waiting for a timeot. Other simlations needed: In addition to the test configrations discssed, simlations wold be needed to test the effect of other interfering sorces like microwave ovens etc. As already indicated previosly also the transfer delay distribtion, or at least the maximm delay a station enconters shold be looked at. This will help also in optimizing the backoff algorithms for CSMNCA and the lost packet recovery retry procedre. Conclsion: It was conclded that the interference robstness and medim sharing behavior are the most important characteristics of a wireless protocol. Based on this a global analyses shows that a distribted "coordination fnction" is best sited to achieve atomatic sharing of the available medim bandwidth, withot additional coordination overhead. To improve the robstness of the protocol a MAC level recovery mechanism is very effective. In addition it is shown that CSMNCA as sed by W A VELAN does have a very high throghpt efficiency paired with a low transfer delay, which is similar than the performance of 82.3 CSMNCD networks. Wavelan CSMNCA, CSMNCA + Ack and the modified Ken Biba 4way LBT protocol are compared on several aspects. A PeerToPeer Bffered Load test methodology is proposed and allows protocols to be compared to several relevant characteristics. The relevant comparison methodologies are explained and illstrated with several simlation otpt charts. A ClientServer test configration is sggested to evalate the net performance for an application. Frther work will be needed to compare the different protocols, and this will be sbject for ftre contribtions. Printed: MAY 8,1992 Page 17 By: Wim Diepstralen

18 Doc: IEEE P S1 WIRELESS MAC PROTOCOL COMPARISON * What are the most importt Protocol characteristics of a Wireless MAC * Global selection of the type of protocol most sitable to meet the most important characteristics. * * * * Comparing Distribted Access Protocols Defme the type of simlations needed Explain the test configrations Describe the comparison methodology * Discss the performance reslts Printed: MAY 8,1992 Page 18 By: Wim Diepstraten

19 Doc: IEEE P S1 What are the most important characteristics of a Wireless MAC Protocol * Protocol characteristics needed for data Brsty traffic reqires low response times Need high average throghpt Can tolerate large transfer delay variations * Services sed for data Connectionless service most commonly sed by todays Network Operating Systems. Isochronos not considered dring comparison Printed: MAY 8,1992 Page 19 By: Wim Diepstraten

20 Doc: IEEE P S1 * * * MAC protocol shold work in single channel environment Environment is INTERFERENCE limited Most important characteristics of a MAC protocol is then: Robstness for interference Medim Sharing Characteristics * Select type of Coordination fnction Centralized Distribted Not able to manage cochannel interference sitation efficientl y Can not manage mltiple ESA interference Need bandwidth for "coordination" Long response times Short response times Can resolve cochannel interference efficiently by recovery strategy on MAC level Atomatic Sharing withot added complexity Printed: MAY 8,1992 Pnge 2 By: Wim Diepstrnten

21 Doc: IEEE P * Main MAC characteristics reqired Access procedre with low "Collision Probability" MAC level recovery * Distribted protocol comparison W A VELAN CSMACA CSMACA + Ack 4WAY LBT (Ken Biba) It is not the intention to come to a conclsion which protocol is best * W A V ELAN CSMACA perfonnance High throghpt efficiency similar to 82.3 CSMACD Low transfer delay Very stable for high loads Printed: MAY Page 2 1 By: Wim Diepstraten

22 Doc: IEEE P * Simlation environment Simlation model explained in Doc IEEE Graphical post processing with SigmaPlot 4.1 * Simlations performed Peerto?er tests Throghpt and Transfer Delay verss Bffered Load Throghpt and Transfer Delay verss nmber of stations. Throghpt as fnction of the PHY preamble length. Medim sharing behavior (at high load). Look at individal station behavior. o Mltiple Networks o Mlti Floor o "Hidden Stations" Client Server Throghpt in a high load Novell environment (Perform3 test) Individal station Printed: MAY Page 22 By: Wim Diepstraten

23 Doc: IEEE P S1 Traffic generation model:!=======================! !xxxxxxx! Random delay Access delay Xmit delay Random delay ++ Access delay U sed to control the "bffered load" dring the test, it is generated randomly with a fixed minimm time of in this case 1 msec. This represents the minimm processing time that is needed to generate a new packet by a station. This will be the time reqired for the MAC to gain access to the medim. xx Xmit delay This is the time reqired to transmit the total packet, at a given raw bitrate. Bffered Load station = l(a verage Random + Xmit) * Average Packet length Total Bffered Load will be the bffered load per station times the nmber of acti ve stations. Printed: MAY Page 23 By: Wim Diepstraten

24 Doc: IEEE PB2.1192S1 * Reslts: CSMACA is very efficient All protocols share medim graceflly 4WAY LBT is highly effected by PHY preamble length Lost bandwidth de to lost "Data" in CSMACA + Ack does cost less than the lost "Data" prevention overhead needed in the 4WAY LBT protocol 4WAY LBT performs better for hidden stations There is still lost "Data" in 4WAY LBT protocol * Conclsion Distribted access protocols can efficiently operate in a Wireless environment They have sperior characteristics for interference robstness and medim sharing. Printed: MAY 8,1992 Page 24 By: Wim Dicpstraten

25 May 1992 Doc: IEEE P S1,.. <1> Throghpt verss Bffered Load Crve 1% Long Packets (1 OBB Bytes) THROUGHPUT : Xfer DELAY :: WAVELAN CSMACA lpersistent CSMA ALOHA V' , >.. 14 en "' :J Q. 1..c CJl :J 8 I..c f 6 4 V f I _"'_..., j V j ' ' 1 '' :...: ::.. I 7V' V' t V' r _ r r _ 2 2 o o VI 14 E '' > I... " VI a 6 4 c I... f Fig: 1 Bffered Load (KByte) I COMPARE ALOHA, CSMA and WAVELAN CSMACA WAVELAN CSMACA Throghpt is 87% of the 2 Mbps raw bit rate. WAVELAN CSMACA is slable al high loads. The Delay is only calclated for those packets that gel throgh. Figres inclde the MAC overhead. Load generated by 7 Stations. Shmission Pae 25 Wim Dienstraten

26 May 1992 Doc: IEEE P Throghpt verss Bffered Load Crve,......, >. CD 1% Long Packets (188 Bytes) THROUGHPUT : Xfer DELAY:: WAVELAN CSMACA 1persistent CSMA ALOHA < '' 12 +' :J... : CJ'I :J 1 8. : I 6 4 I V V.,'. 1 r :l L:! V r! o i t i i:v = :3... c <7 f <7.. I I o o Fig: la Bffered Load (KByte) COMPARE ALOHA, ' CSMA and WAVELAN CSMACA r.. (fj 14 E '' >. 12 E 1 \.., (fj c: I WAVELAN CSMACA Throghpt is 87% of the 2 Mbps raw bit rate. WAVELAN CSMACA is stable at high loads. The Delay is only calclated for those packets that get throgh. Figres inclde the MAC overllead. Load generated by 7 Stations. Sbmission Page 26 Wim Diepstraten

27 May 1992 Doc: IEEE P Throghpt verss Bffered Load Crve 1% Long Packets (188 Bytes),.,. <1l. >, m '' +' :J...!: > :J L...c I THROUGHPUT: WAVE LAN CSMACA CSMACA + Ack 4WAY LBT,j g f t,. V 1 :: f:::.:.. :::::: '''''' I.7 f...f.. V I,lit",..,. i :=i ;::..::::..., v Xfer DELAY:: 'V I I E '' >. 24 <1l o 2 <1l 16 c o L. 12 I 8 4 Fig: 2 o o Bffered Load (KByte) COMPARE CSMACA, CSMACA + Ack and 4WAY LET WAVE LAN CSMACA Throghpt is 87% of the 2 Mbps raw bit rate. WAVELAN CSMACA is stable at high loads. The Delay is only calclated for those packets that get throgh. Figres inclde the MAC overhead. Load generated by 7 Stations. Sbmission Wim Diepstraten

28 May 1992 Doc: IEEE P Throghpt verss Bffered Load Crve r... Cl) >. m '" :J....c Q'1 :J '.c f % Long Packets (188 Bytes) THROUGHPUT: WAVELAN CSMACA CSMACA + Ack 4WAY LBT,.li.I L 1 V b J ' t' 1! " a..::i '" V i' a "._ Xfer DELAY:: f V' f I o ,... \11 28 E '" >. 24 E <U o 2 L. Cl) '+ \11 16 c a L. 12 f 8 4 o Fig: 2a Bffered Load (KByte) COMPARE CSMACA, CSMACA + Ack and 4WAY LET WAVE LAN CSMACA Throghpt is 87% of the 2 Mbps raw bit rate. WAVELAN CSMACA is stable at high loads. The Delay is only calclated for those packets that get throgh. Figres inclde the MAC overhead. Load generated by 7 Stations. Shmission Pae 28 wim Dieostraten

29 May 1992 Doc: IEEE P S1 Throghpt verss Bffered Load Crve 2: % Short Packets (64 Bytes), 4% Long Packets (576 Bytes) THROUGHPUT : Xfer DELAY:: WAVE LAN CSMACA SMACA + Ack 4WAY LBT \ r... '' 16 >.. co 14 ::"c ''" 12 +' :::J Q.. 1 > :::J 8 f 6 4 Fig: 3 'I J If ) ",v 1. ",\1 2.v f I \1 v _\1 nfo 1 I::e : 2 o o Bffered Load (KByte) 16.r... (I 14 E ''" > ClJ o 1... ClJ " (I 8 c: o... f 6 COMPARE CSMACA, CSMACA + Ack and 4WAY LET 4 The Delay is only calclated for those packets that get throgh. For CSMACA the lost packets are not recovered so it is not inclded in the delay figre. Figres inclde the MAC overhead. Sbmission Page 29 Wim Diepstraten

30 May 1992 Doc: IEEE P r.. OJ Performance verss Nmber of Stations Crve 6% Short Packets (64 Bytes), 4% Long Packets (576 Bytes) THROUGHPUT : Transfer Delay:: WAVELAN CSMACA CSMACA + Ack 4WAY LBT "l " > 1 4 en '" 12 +' ::J Q. L 1 (J'I ::J I... 8 L f 6 Fig: 4 1 r f r I.( : 2 cp :::. ( o o , ".,,. )j i. _ 1 < p j (., Nmber of Stations ' ,... Vl 14 E... > 12 OJ 1 I... OJ COMPARE CSMACA, CSMACA + Ack and 4WAY LET Vl c: a I... f Figres inclde the MAC overhead. Throghpt remains stable for many simltaneos stations accessing the medim. For CSMA CA lost pac kets are not recovered at MAC level. Sbmission Page 3 Wim Diepstraten

31 May 1992 Doc: IEEE P S1 Performance verss Nmber of Stations Crve 6% Short Packe ts (64 Bytes). 4% Long Packets (576 Bytes) THROUGHPUT : % Data errors:: W A VELAN CSMACA CSMACA + Ack "7 4WAY LBT ' >, 14 co "" 12 +' ::l.. 1 L (j1 ::l 8 L. L f 6 V I " ;1 f? j ' ' 11 7 I 4 7" '.,? 2 2,.., o o 2, ; Ul +'.::,t.... L. ' ' w Fig: 4a Nmber of Stations COMPARE CSMACA, CSMACA + Ack and 4WAY LET Only lost "Data" packets are conted. The 4 WAY LET protocol does have very low "Data" error probability, becase access contention is resolved by the RTS. CTS handshaking.. Althogh CSMA CA + Ack has to recover extra lost "Da ta" pac kets, the throghpt is still significant higher than in the 4WAY LET protocol. Figres inclde the fac overhead. Sbmission Page 31 Wim Diepstraten

32 May 1992 Doc: IEEE P Performance verss PHY preamble length rve r., +J >, ill 6% Short Packets (64 Bytes), 4% Long Packets (576 Bytes) THROUGHPUT : Transfer Delay:: WAVELAN CSMACA CSMACA + Ack 4WAY LBT <: '' 12 +J :::J Cl. '= CJ) :::J 1 8 ' '= f 6 "... i " (D, r "'. "'...,.. r... 1 < r....1 r I,... ', ==( 1'= ( p( b o o Fig: 5 Preamble length (bits) E '' >, 12 ClJ 1... COMPARE CSMACA, CSMACA + Ack and 4WAY LET ClJ c... f Figres inclde the MAC overhead. Clear efficiency degradation for long PHY training times. The higher the modlation rate, the longer the PHY training time in nmber of bits. Test configration is: 7 stations at high load. Sbmission Page 32 Wim Diepstraten

33 May 1992 Doc: IEEE P VUS.L.LtJ" V ''.I. U l',,,, V.L no. ''tj..l.".lv.l.l PeerToPeer Test configration 4% Short Packets (64 Byte) 6% Long Packets (519 Bytes)....., >. 25 m Y: '...., ::J a.. '= CJl ::J " '= t THROUGHPUT : WAVELAN CSMACA CSMACA + Ack 4WAY LBT Tc tal Bo lh Net wor ks P, 1 ' 'C1 1'7 p v:::... e_ ><. '>< <J... loa I.,=... ' h I ' p=ij h=po_ p D_,. 'V... >, 1'"' Xfer DELAY:: e 1 'C1 e 'C1 p p e i" _ ppp Ul 25...s >. a 2 '+ 15 a I... t 1 5 Fig 6 o o Distance Between NWs (m) Compare Mltiple Network Performance Note that when two networks flly overlap then Bffered Load is generated which is eqivalent to 14 stations. Sharing similar for all 3 protocols. Sbmission Note that environment is considered homegeneos withot any obstrction by walls e t c. Page 33 Wim Diepstraten

34 May 1992 Doc: IEEE P v Vl Lf) "" () I... v a. +' v.::.: Throghpt verss Network separation PeerToPeer Test configration 4% Short Packets (64 Byte) 6% Long Packets (512 Bytes) v Lf) "" () I... v a. +' V.::.: Distance Between NWs (m) FIG. 7: v Lf) '" l v a. +'..Y... Distance Between NWs (m) WAY LBT Sbmission Page 34 Wim Diepstraten

35 May 1992 Doc: IEEE P S CIJ 5 If) 45 ' ' CIJ U o CIJ 5 If) 45 ' : 3 ' CIJ a ' CIJ CIJ 5 If) 45 ' ' CIJ a ' CIJ..:: Throghpt verss Network separation PeerToPeer Test configration 4% Short Packets (64 Byte) 6% Long Packets (512 Bytes) Q::. W,ViLNM. lt:., ji ;;t,ii ; ",. " I;:. ;i ;:::...: I,...:. :::..:. "'.,.,," '...,, 'C1 ' 6 '' '' I,,.'.', y',"; ' I ' 6 r. ; i". ",.. \: "." JII! 6 2 db Between NW's o o Distance Between NWs (m) 2 db Between NW's WAY LBT Distance Between NWs (m) 1 5 FIG 7a: Sbmission Page 35 Wim Diepstraten

36 May 1992 Doc: IEEE P S1 +' (I) (I) +' 15 Cl.. U1 f 1 ::: Individa l Station performance PeerToPeer test configration All stations NWl 1 'V 2 T 3 o Distance between NWs (m) 25 I All stations NW 1 ". 2 o " 1'] 15 c o Cl.. () 1 l f U FIG. \ \. '.' ' \ I, (!i' " :. \ 6 ' 'i T,' '., " "' 6 o, ' i ' IY., j _.. ' '.._ 5 ' I T" :.' ". ::""'" ==.: o I j I' '. o 1 2 3e FIG. 8: Distance betwee n NWs (m) 4vVA Y LET PROTOCOL Sbmission Page 36 Wim Diepstraten

37 May 1992 Doc: IEEE P Individa l Station p e rformance +' U1 o U1 +'.::.:. o Q (fj PeerToPeer test configration A l' stations NW1 1 \l 2.." I 2 db Between Nw'1 Distance between NWs (m) 3 "U. U1 "U U1 +'.::.:. Q (fj f U All stations NW1 I f:,. I. \ \ I..." ". \ I FIG. 8a: Distance between NWs (m) 4 vva Y LET PROTOCOL Sbmission Page 37 Wim Diepstraten 6

38 (j) 2... >. CD :::.:::... :::J Q...c ' :::J.....c l 15 1 SO Doc: IEEE P Throghpt verss Network separation ClientServer test configration 4% Short Packets (64 Byte) 6% Long Packets (576 Bytes) p THROUGHPUT: WAVELAN CSMACA CSMACA + Ack 4WAY LBT Tc tal Eo th Net wor ks,. rv "... ',. " " 1_ '7'" '". pp " v,.' v ie, f", Ia o "'"::". p.. v,, pop'"' Xfer DELAY:: e <:l '.:: V :., I!e e_ =e : o o ,. (I) 25 >. o 2... (j) ' o 1 5 Fig 9 Distance Between NWs (m) Compare Mltiple Network Performance Note that when two networks flly overlap then Bffered Load is generated which is eqivalent to 14 stations. ONL'I STF1TIOtJ To S 2UE'R.. DELR'j :;I LV;{ Sharing similar for all 3 proocols. Sbmission Note that environment is considered homegoneos withot any obstrction by walls etc. page 38,\Vim Diepstraten

39 May 1992 Doc: IEEE P S1 Throghpt verss Network separation ClientServer test configration 4% Short Packets (64 Byte) 6% Long Packets (512 Bytes) 3 r'''''rr. 25 V1 L.{) '... 2 ' 15 a.. V1 1 +' '" a 5 Q Distance Between NWs (m) 3 rriri' II!r'IT I 25 r CSMACA + Ack If) '... 2 t 15 r a.. C) +' '" a Q 1 o I i I I I I i G Distance Between NWs (m) 3 r'i'i'iri'i'i''i' 25 r 4WAY LET '... 2 t ' a.. V1 +' '" a Q o I i i i o Sbmission Page 39 W im Diepstraten

40 May 1992 Doc: IEEE P JOO 25 " 2 15 ' Q +'.;,,: o Q 1 Throghpt verss Network separation ClientServer test configration 4% Short Packets (64 Byte) 6% Long Packets (512 Bytes) WAVELAN CSMACA '. ('V\ : '.1'... 'V..... " ", I... \.. \'" I( \....! a t::;... ' :i '.:. ":<;=b :1 I. (,::!,, '". '.. ;1'>.. '', v ">.,. I... '.... T; D', v Q I. \ I,, \. Q' " < g?:.. I;l ' ':' I. ",..' ' '....., 5 Between NW's! If) " ' Q +'.;,,: CL Distance Between NWs (m) JOO 25 CSMACA + Ack db Between NW'sl I Lfl " L Q Ul +'.;,,: Q JOO WAY LBT Distance Between NWs (m) FIG. loa: Sbmission doo Page 4 wim Diepstraten