A Real-time Medium Access Control Protocol for Ad hoc Wireless

A d h o e

Rusty O. Baldwin Nathaniel J. Davis IV Scott E Midkiff baMwinr@ vt. edu ndavis @ vt. edu midkiff@ vt. edu

Center for Wireless Telecommunications, Bradley Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State University,

Blacksburg, VA

We develop and analyze a simple, elegant medium access control (MAC) protocol for use inn transmitting real-time data in point to point ad hoc wireless local area networks (WLANs). Our enhancement of IEEE 802.11, real-time MAC (RT-MAC), achieves dramatic reductions in mean delay, missed deadlines, and packet collisions by selectively discarding packets' and sharing station state information. For example, in a 50 station network with a normalized offered load of O. 7, mean delay is reduced from more than 14 seconds to less than 45 ms, late packets are reduced from 76% to less than 1%, and packet collisions are reduced from 36% to less than 1%. Regression models are developed from simulation data to describe network behavior in terms of throughput, mean delay, ratio of late packets, and ratio of collisions. Stations using RT-MAC are interoperable with stations using IEEE 802.11.

I. In troduc t ion

Communications pervades modern society. From broadcast radio and television, to the exchange of information via two- way radios, telephones, cellular phones and pagers, to the global internet, communications pervades every aspect of our lives. It has also had an enormous impact on the industrial and manufacturing industries. Production lines and industrial control systems rely more and more on computers, often several, to control manufacturing processes and robotic assembly systems. These computer systems in turn require timely (i.e., real-time) information via communication networks to coor- dinate their actions. Timely information is vital to the suc- cess of military operations as well. Modern warfare requires extensive communication between the military services (e.g., the Army, Air Force, Navy, etc.) as well as between units in those services. Moreover, military personnel in the field need to communicate with their weapon systems which are often controlled remotely. To provide this timely information, new protocols must be developed which support the transmission of real-time data such as voice, video, and automatic control information within a given amount of time. This paper describes a novel protocol used to support such data transmission in the context of a point to point ad hoc wireless local area network (WLAN).

From the earliest wireless LANs such as ALOHA [1], research into wireless LANs has continued uninterrupted. Early research identified fundamental principles and analysis tech- niques [2], [16], [18], [19], [201, [21], [221, [33] which are still applicable, as well as fundamental problems that are still encountered [31], [32]. Much, if not most, of past research has been focused on increasing throughput and reducing mean delay. More recently, a measure of attention has turned to the area of real-time WLANs where individual packet delivery times are the foremost concern [7], [10], [24], [26], [29], [36]. Excellent surveys of work in real-time LANs can be found in [23], [271.

Most real-time systems are specialized; designed and built to satisfy a unique requirement. As such, these systems are

20

typically expensive and not easily transferred to other application areas. Given the increasing demand for real-time systems, especially in the areas of voice and video data, a low-cost solution to real-time communications is highly de- sirable. IEEE 802.11 is a recent (1997) standard developed for WLANs [14]. It has capabilities that can be exploited to provide real-time service. A standards-based solution offers the potential for a low-cost implementation of an effective real- time system. Additionally, 1EEE 802.11 supports probabilistic access to the medium which can be used to support a wide range of application areas.

The remainder of this paper is organized as follows. Sec- tion II briefly describes IEEE 802.11. Section III describes the modified IEEE 802.11 protocol, RT-MAC. Section IV ex- plains the simulation model, methodology, and assumptions used. Section V presents an evaluation of RT-MAC and regression models derived from simulation data. Conclusions and future work are discussed in Section VI.

II. IEEE 802.11

The IEEE 802.11 standard defines both a multiple access control (MAC) protocol and physical layer implementations. At the MAC layer, IEEE 802.11 supports both infrastructure and ad hoc networks. In an infrastructure network, stations are granted access to the medium by a station known as the point coordinator (PC) and use the Point Coordination Function (PCF) for packet transmission. Due to the overhead in this cen- tralized access scheme, PCF have been regarded as unsuitable for real-time data [29], [34]. In an ad hoc network, stations "compete, for access to the medium using the Distributed Co- ordination Function (DCF). Both infrastructure and the ad hoc networks ultimately use DCF for medium access. The DCF prioritizes access to the medium by specifying a time interval between frames known as the inter-frame space (IFS). By definition, during an IFS the medium is idle. The different types of IFSs, along with the backoff mechanism described below, allows a station to determine whether it may transmit. It is

Mobile Computing and Communications Review, Volume 3, Number 2

Immediate access when medium is free > DIFS

i ~ N e x t Packet Slot Time

iSelect slot and decrement backoff as long as medium is idle

Figure 1: IEEE 802.11 Basic Access Method

also known as the basic access method. There are four types of IFSs: Short IFS (SIFS), PCF IFS

(PIFS), DCF IFS (DIFS), and Extended IFS (EIFS). EIFS, which is the longest IFS in terms of time, is used when bit errors introduced by the physical medium cannot be corrected by the radio receiver. Transmission after SIFS, the shortest IFS, is reserved for the PC to send any type of frame required or for other stations to begin transmission of an acknowledg- ment (ACK) frame, a clear-to-send (CTS) frame, to respond to polling by the PC, or to send a fragmented MAC protocol data unit (MPDU). Similarly, access after PIFS is reserved for stations to begin transmission of PCF traffic. After DIFS, if a station determines that the medium is idle, it may transmit a pending packet. If the medium is not idle after DIFS, a backoff timer is set by selecting a random integer (hereafter known as the backoff value (BV)) from a uniform distribution over the interval [0, CW-1 ], where CW is the width (in slots) of the con- tention window. This BV is the number of idle slots the station must wait until it is allowed to transmit. For every idle slot de- tected (after a DIFS), the timer is decremented by one. I f the medium becomes busy prior to the timer expiring, the timer is frozen until the next DIFS, upon which the timer decrements again. Upon expiring, the station transmits its packet. If there is a collision, CW is doubled until it reaches a predefined maximum value, CWmax. Upon a successful transmission, CW is reset to the default minimum value of CWmin. Figure 1 [14] shows the structure of the basic access method.

At the physical layer, IEEE 802.11 uses carrier sensing to detect an active channel. It supports a 1 Mbps and a 2 Mbps data rate. Recently, proposals to support data rates up to 30 Mbps as well as data encoding to enhance the coverage area has been introduced [9]. The interested reader is encouraged to consult [8], [12], [14], [34], and [35] for more detailed information on IEEE 802.11.

its deadline; collision avoidance is achieved by deferring BV decrements while the medium is busy and by doubling CW upon transmission failure as described in Section II.

RT-MAC uses two additional pieces of information to achieve its result: a transmission deadline and the transmitting station's next BV. When a real-time packet is submitted for transmission, a transmission deadline (i.e., when transmission of the packet must begin by) is associated with the packet. This value is only needed until the packet is either successfully transmitted or discarded and therefore does not become part of the packet itself. When the packet is removed from the queue for transmission the next BV to be used by the transmitting station (i.e., the number of idle slots the station will wait prior to transmitting again) is placed in the packet header. Stations that hear the transmission will use this BV to avoid selecting the same BV for their backoff timer. This will be described in detail in Section III.B. Note that BV is NOT the value of CW (cf., Section II), but rather the result of the random selection of a BV from the range of [0, CW-1].

IILA. T r a n s m i s s i o n C o n t r o l

The transmission deadline of a packet is examined at three key points to determine whether to discard the packet. By discarding a packet as soon as possible after determining that its deadline has been exceeded, transmission queue throughput is increased and as a result, the likelihood that other packets in the queue will meet their deadlines is increased. The examina- tion points (described below) were chosen because they follow unpredictable delays that a packet suffers prior to transmission.

A packet is first examined when it is removed from the transmission queue in preparation for transmission. If the packet has already exceeded its transmission deadline, it is discarded and the next eligible packet in the queue (if any) is selected. At this point, the station may need to wait for the backoff timer to expire. During this time, other stations could possibly transmit. After the backofftimer expires, the packet is examined again. I f the packet deadline has been exceeded the packet is discarded, otherwise, it is transmitted. Assuming the transmission is successful, the next eligible packet is selected and the process repeats. If the transmission is not successful (that is, no acknowledgement is received), the packet deadline is again examined and the packet is discarded if the deadline has been exceeded. If the deadline has not yet been exceeded, the packet is submitted for retransmission. Using this transmission control (TC) algorithm, a packet that is successfully received will never be late. The TC algorithm is summarized in Figure 2.

III. RT-MAC Description III .B. E n h a n c e d C o l l i s i o n A v o i d a n c e

Two major factors that impact the ability of a real-time WLAN to meet packet deadlines are the transmission of packets that have already missed their deadlines and packet collisions. Packets that have missed their deadlines are assumed to be unusable by the receiving station so transmitting them consti- tutes a double failure. The first failure is the missed deadline itself, the other is the wasted channel capacity that could have been used to transmit a usable packet. IEEE 802.11 does not have any means of detecting whether a packet has exceeded

The enhanced collision avoidance (ECA) algorithm has two components. First, rather than employ a static value for CW, CW is set to be eight times the number of stations in the network (i.e., CW = 8N). If a static CW value is used, the probability of two or more stations choosing the same BV increases with N. By making CW a function of N, this probability does not increase. N is assumed to be estimated either by tracking the number of unique station addresses that have transmitted over the last z seconds where z is a suitable value, or by

Mobile Computing and Communications Review, Volume 3, Number 2 ~.. 21

F

Discard

Packet

+

t Remove Packet

Prom Queue

(yes)

(yes)

Discard

Packet

(no)

RT = Rea l - t ime V--- RBV = Received Backof t Value

C B V = Currea t Backof f Value

Figure 3: Enhanced Collision Avoidance (ECA) Algorithm

the receiving stations current BV. If a suitable value cannot be found, the range of values will be doubled (i.e., [0, 2CBV-1]) until a suitable value can be found. Figure 3 summarizes the second component of the ECA algorithm.

iV. S i m u l a t i o n M o d e l

Figure 2: Transmission Control (TC) Algorithm

a method such as the one described in [8] where N is estimated as a function of channel load. The value of 8 is loosely based on the CW equation used in [8]. In [8], this value is determined based on packet transmission time and by estimating channel utilization over an arbitrary observation period. Rather than estimating these parameters, we use a fixed value of 8 and rely on the second component of our algorithm (described below) to resolve any BV conflicts that may still occur. Several schemes that dynamically alter the value of CW have been proposed and can be found in [7], [8], and [10].

The Second component of the ECA algorithm involves ad- vertising the transmitting station's next BV and tracking the BVs of other stations in the network. Prior to transmitting a packet, a station will select its BV from the range of [0, CW- 1], excluding BVs that are known to be in use. This selected BV will be the one used following the current transmission. It will also be placed in the packet header and transmitted along with the packet. Stations that hear the transmission will place the BV in a table of BVs "in use". During idle slots, a station will decrement its own BV (as in IEEE 802.11) as well as every BV in its table of BVs.

A station may receive a packet that indicates the sending station has chosen the same BV as the receiving station. This could occur due to new stations joining the network or due to BVs not being received because of collisions or bit errors. In such cases, the receiving station chooses another BV since a collision will certainly occur (assuming both stations have a packet to transmit). To prevent a station that must choose a new BV from being unduly penalized, the new BV is chosen (if possible) from the range of [0, CBV-1] where CBV is

22

To investigate the performance of our protocol, a simulation model of the DCF (cf., Section II) of IEEE 802.11 was developed and implemented using the network simulation tool OP- NET version 3.5A [28]. The model was developed using the Specification and Description Language (SDL) [ 15] specification of IEEE 802.11 that is found in Annex C of the standard. The model was validated against the results obtained in [8]. Because [8] was based on an earlier draft, several of that paper's parameters did not match those in the latest IEEE 802.11 standard. For the validation, we changed the values of these parameters in our model to match [8]. For a more detailed dis- cussion of the model construction and validation readers are encouraged to consult [4].

The network under investigation is a fully meshed ad hoc network of 5 to 50 homogeneous stations, which is consistent with other studies [7], [8], [10]. The physical layer is assumed to be a Direct Sequence Spread Spectrum (DSSS) transmission system. The remainder of this section describes the simulation parameters, factors, and response variables.

IV.A. S i m u l a t i o n P a r a m e t e r s

Several parameters are used to control the simulations. The simulation model can vary over 30 different parameters, most of which are IEEE 802.11 parameters. The most significant parameters are listed in Table 1. Any other parameters not listed are either not implemented in the model or are set to the default value for a DSSS system.

IV.B. S i m u l a t i o n Factors

Factors are parameters that are varied during the simulation such that they significantly impact system performance when altered [17]. The set of factors used include the number of stations, N, the total offered load, G, the channel model, E,


Tabte 1: Simulation Parameters Parameter Value

-X~iT~ics

Channel Bit Rate CWmin CWmax Slot Time SIFS DIFS ACK Length PHY Header ACK Timeout RxTx _Turnaround_Time

1 Mbps 31"* 1023"* 20 # s 50 # s Calculated at runtime Calculated at runtime Calculated at runtime Calculated at runtime 5 ~ s

** IEEE 80211 oaly

Table 2: Traffic Models Factor Value

tnterarrival Distribution Constant Deadline Distribution Constant

(same as interarrival time) Packet Size (bytes) 83 Discarded Packets Resubmitted 0% Interarrival Distribution Deadline Distribution

Packet Size (bytes) Discarded Packets Resubraitted

Poisson Truncated Normal Mean = 380 ms, Min=21 ms Max= 1 s 775 50%

Table 3: Simulation Factors [ Factor 'Values

[ Number of Stations ( N ) 5, I0,20.30,40,50 Total Offered Load (G) 0.3 0.5,0.7,0.9 Channel Model (/5) 0 (Idesd), I (Bursty) Protocol [EEE 802.11 RT-MAC

VLC. R e s p o n s e V a r i a b l e s

Four response variables are of interest. Of primary interest, given the real-time emphasis, is the missed deadline ratio, F. The missed deadline ratio is number of packets that exceed their deadline over the number of packets removed from the queue for transmission (cf., Figure 2). A discarded packet (due to exceeding the transmission attempt count or due to the TC algorithm) is deemed to have exceeded its deadline. A second response variable is the collision ratio, C. The collision ratio is the number of packet collisions over the number of transmission attempts. The third response variable is the aggregate system throughput, S. Finally, the fourth response variable is the mean delay of successfully received packets, D. The sub- scripts Std and RT, when used with the response variables, mean the IEEE 802.11 and RT-MAC protocols, respectively.

where E = 0 means an ideal channel and E = 1 means an errored channel, and the protocol in use (IEEE 802.11 or RT-MAC). We investigate two types of traffic: telemetry and avionics. The telemetry traffic model is representative of traffic on a MIL-STD-1553B serial data bus [3], while the avionics traffic model is representative of the requirements for traffic on the avionics bus on the Boeing 777 [11].

The telemetry model has short packets and constant interarrival times such as can be found in a cyclic real-time system. It is intended to be a "worst-case" traffic model in the sense that short packets induce greater overhead as well as more oppor- tunity for collisions. The transmission deadline for the packets is equal to the interarrival time. The total offered load, G, is varied by adjusting the interarrival time appropriately. Pack- ets discarded by the TC algorithm are assumed useless to the receiver and are not resubmitted for transmission.

The avionics traffic model has moderately sized packets which arrive according to a Poisson process. Deadlines are normally distributed. The total offered load, G, is varied by adjusting the mean interarrival time appropriately. In this traffic model, some of the packets discarded by the TC algorithm are assumed to still be useful if transmitted. Therefore, 50% of discarded packets are randomly resubmitted to the transmission queue. Table 2 provides details of the two traffic models.

The other factors that were varied and their values are shown in Table 3. The channel model used in the simulations was rather ideal (E = 0) or an errored channel (E = 1) using a two- state bursty error model as in [51, [6], and [13]. The errored channel is in a "good" state for an exponentially distributed amount of time with mean 5.0s. In the good state, there are no bit errors. It is in a "bad" state for an exponentially distributed amount of time with mean 0.1 s. In the bad state, the probability of a bit error is 0.8. The error probability and state times are based on the observed behavior of real WLANs [5], [131. The average BER varies with G but is typically quite poor, with B E R ~ 2 x 10 .2 .

V. Performance Analysis

A full factorial experiment with replications was chosen for this effort. At least five replications were needed to obtain suitable confidence intervals for the response variables [17], [25]. A simulation run was terminated when the confidence interval widths of all four response variables were within a given percentage of their mean values [17] or after a simulation time of 100s (telemetry traffic model) or 1000s (avionics traffic model). This termination criteria helped to ensure that the system was in steady-state. The throughput and mean delay confidence interval widths were required to be less than or equal to 2% of their mean value and the missed deadline and collision ratio confidence interval widths were required to be less than or equal to 10% of their mean value. Data collection began after 5s of simulation time: Using the factors in Tables 2 and 3 and performing five replications results in a total of 960 simulation runs. Individual simulations runs were performed on several machines simultaneously. They took the equivalent of approximately 8000 CPU hours on a Sun UltraSparc 2 with 256 MB of RAM a 170 MHz processor running Solaris 2.5.

V.A. S i m u l a t i o n Resul t s

We found dramatic and noteworthy performance improvements using RT-MAC. In general, for both the telemetry and avionics traffic models, RT-MAC stabilized the behavior of the response variables whereas IEEE 802.11 was characterized by asymptotically increasing mean delay, collisions, and missed deadlines beyond a certain total offered load. Figure 4 is a case in point.

Figure 4 shows Dstd and DnT using the telemetry traffic model and an ideal channel. Dstd increases asymptotically as G increases from 0.3 to 0.5. DnT actually decreases as G increases. This behavior of RT-MAC can be attributed to the TC algorithm. Since late packets are discarded rather than transmitted, the throughput of packets in the transmission queue is

Mobile Computing and Communications Review, Volume 3, Number 2 ....... .... 23

100 -

"g" 10

0.1-

e~ ,~ 0.01-

I E E E 8 0 2 . 1 1 . . . . . . . . . R T - M A C

N : 550 ......... :::::::::::::::::::::::::::: N = 40 - - ~fi~'~-2222222222222252222222222222222222 5' = 30 - - %",', ...................................

N : 2 0 - x g V / N = 10 - - : , z Z 2 v L ¢ , . . . . . . . . . . } - - 2 7 , - . . . . . . . . . . . . . . . . . . . .

.%',,, ," ," / / / N = 40 . . " . ; -> . ' . - " / / / ~ \ , : ao

,',,;'," .'," N = 2o . . ' . ? / . . ' . " N = lo

I I . . . . I t - - I -- I I 0.3 0.4 0,5 0.6 0.7 0.8 0.9

Total Offered Load (G)

Figure 4: Mean Delay -Telemetry Traffic Model

0 . 4 -

0.3-

g

0 . 2 -

0.1-

I E E E 8 0 2 . 1 1 . . . . . .

G = 0.5,0.7,0.9 . ~ ' ~ " .* ' "

, " " ~ ~ ' ' " G = 0.3

G : 0.30.9

0 - I T If ~ t I0 20 30 40 50

Stations (N)

Figure 5: Collision Ratio - Telemetry Traffic Model

much greater than that of IEEE 802.11 and consequently the mean delay decreases. In the simulation model, the transmission queue was limited to 200 packets. Using IEEE 802.11, as G increased from 0.5 to 0.9, the number of packets blocked due to a full transmission queue increased from 10% to 55% of arriving packets. For RT-MAC, the queue length never exceeded 10 packets. For the errored channel, the IEEE 802.11 network typically experienced a 10 to 250 ms increase in mean delay. The RT-MAC network had no increase in mean delay since the packets that became late due to retransmissions were discarded by the TC algorithm.

For an ideal channel, Fsta was close to zero for G = 0.3, N < 40 and reached 1.0 for G > 0.5. FRT increased approximately linearly from 0.0 to 0.63 (independent of N) as G increased.

Packet collisions, C, showed similar behavior. Figure 5 compares the packet collision ratios of IEEE 802.11 and RT- MAC. Csta is characterized by rapidly increasing collision ratios as G increases from 0.3 to 0.5 and a maximum collision ratio strongly influenced by N. CRT shows extremely stable behavior. It never exceeds 0.045.

Aggregate throughput, S, for both protocols was generally about 0.3. Usable throughput, U, (the throughput of packets that arrive prior to deadline expiration where U = S(1 - F ) ) was 0.0 for Ustd for any G ~ 0.5 while URT achieved 0.13 to 0.23 for similar network configurations.

24

Table 4: Telemetry Traffic Model Simulation Results R e s p o n s e

Variable

R a n g e o F - R e s p o n s e

Variable Values 90% Confidence Interval

is l e s s than or equal to

T h r o u g h p u t 0.28 _< Sstd < 0.33 _+0.0033 0.28 < SRT < 0.34 ± 0 . 0 0 3 2

Delay (sec) 0.006 < D s t d < 17.0 ± 0 . 0 0 2 7 , D s t d < 1.70 ± O . 0 7 1 , D s t d > 1.70

0 .002 _< D R T _< 0 .043 i 0 . 0 0 0 2 1 M i s s e d Dead- 0.01 < / /S td _< 1.00 ~ 0 . 0 0 3 8

l ine R a t i o 0.0 < F R T _< 0.63 ± 0 . 0 0 5 1 C o l l i s i o n 0.03 < C6'td _< 0.40 : i :0 .0045

Ratio 0.01 ~ CRT ~ 0.04 ~ 0 . 0 0 2 5

• 2 0.1 ~o

g O.Ol

0201-

7, 10 4_

G=0.9

IEEE 8 0 2 , 1 1 . . . . . - . . . . . . . . . . . . . . . . . . . . . . . RT-MAC / G = 0.7

/ G : 0 . 9

/ "

- - ' - ' ' ' - - C=0.7 /

/ '

/ G : 0.5 /

G ~ 0.3

- - I 7 - - I I I 10 20 30 40 50

Stations (N)

Figure 6: Missed Deadline Ratio - Avionics Traffic Model

For the telemetry traffic model, the range of response variable values is shown in Table 4. Note that the minimum and maximum values shown in the tables do not necessarily correspond to the minimum load/station or the maximum load/station network configurations. Using the regression models in Section V.B, the behavior of particular network configurations can be determined. Each of the response variable samples had a slightly different confidence interval width, therefore, the maximum confidence interval width is shown in the table.

Comparable results were obtained for the avionics traffic model. Using this traffic model, IEEE 802.11 performed fairly well for G < 0.5 and N < 20. For other network configurations, behavior in terms of D, F , and C was similar to the telemetry model.

Figure 6 shows Fsta and FRy using the avionics traffic model and an ideal channel. Fsta increased asymptotically as G increased from 0.7 to 0.9 for N < 20. FRT was much more stable, never exceeding 0.14 even in worst-case conditions.

For the avionics traffic model, C and D also showed behavior comparable to the telemetry traffic model. For the avionics traffic model, the range of response variable values is shown in Table 5. As with the telemetry table, the minimum and maximum values shown do not necessarily correspond to the minimum load/station or the maximum load/station network configurations and the maximum confidence interval width is shown. The increased throughput can primarily be attributed to the larger packets sizes.

Other simulation studies were conducted to determine the relative benefit of the different components of RT-MAC. To determine this, we ran simulations (with 5 replications) using a network with an ideal channel, avionics traffic model, and N = 40, G = 0.7. Figure 7 gummafize the results. In terms


Table 5: Avionics Traffic Model Simulation Results Response Range of" Response 90% Confidence Interval

Variable Variable Values is less than or equal to

Throughput 0.29 < SStd ~ 0,76 ~ 0 . 0 0 7 0 0,29 _< ,5'RT < 0 .83 ± 0 . 0 0 3 9

Delay(sec) i 0 .009 < Ds'td < 36.7 ~ 0 . 0 0 2 8 , [_)st d ~ 4.20 -~0,028, DStd > 4.20

0 .009 ~ Dt~T < 0.134 ± 0 , 0 0 0 9 0 Missed Dead- 0.0 < F.vt~i _< 0.99 ± 0 . 0 0 3 8

line Ratio 0.0 < F R T < 0.14 ::t=0.0028 Collision 0.002 < Cstd _< 0.40 ± 0 . 0 0 5 5

Ratio 0 .0008 <_ (TRT <_ 0.03 ~ 0 . 0 0 1 5

s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . "

0.7

0,6

0.5

0.4

0.3

0.2

0.I

0.0 Normalized Mean Delay Missed Deadline [ Collision Ratio

Throughput (S) (sec) (D) Ratio (F) I (C)

[] IEEE ,802.11 0.6421 10.8162 [ 0.5926 [ 0.2079 [ ia rc only 0.6868 o.oa,a o.o,ez [ 0.07z4 [=ECA Only [ 0:6948 , 0:0410 i 0"0079 t 0.0096 121RT-MAC ] 0.691_8 ........ 0,0383 _ 0,0068 ~ 0.0091 .......

Figure 7: RT-MAC Components Study

of F and C, RT-MAC was most effective in minimizing these metrics. In the case of D, RT-MAC was higher than TC Only demonstrating that the CW value (CW= 8N) is not optimal with respect to mean delay.

Similarly, to determine the relative contribution of CW expansion alone versus CW expansion and tracking BVs we ran simulations using a network with an ideal channel, telemetry traffic model, and N = 10, 50; G = 0.7. We found that C was reduced the most when both aspects of the collision avoidance algorithm were active. Both the CW expansion and BV tracking reduced collisions roughly equally when enabled sep- arately.

V.B. R e g r e s s i o n Models

Response variable data was gathered for all combinations of the factors from Table 3 for both traffic models. The statistical significance of the data was then evaluated using the SAS statistical package [30]. The general linear model (GLM) proce- dure was used to evaluate the impact of the simulation factors and their interactions with the response variables. Data trans- formations were required for DStd, Fstd, and Cstd due to the wide range of values obtained. After trying several transfor- mations, it was determined that the power transformation (with a = 0.05) resulted in the best regression model for Dstd and the arcsin transformation was best for Fsta and Cstd.

For each response variable in both traffic models, a regression model was developed and an R 2 value was calculated. Statistically, R 2 measures how much variation in the response variable can be accounted for by the factors used to develop the

Table 6: Telemetry Traffic Model Regression Models

Variable Regression Model

Throughput S$td : 5 .292 x 10 - a G a N -- 6 . t 5 6 × I O - a G U N --0.204(72 + 0 . 3 0 5 G + 0.222

SRT : - -2 .215 x 10 ~ G N ~ + 1.782 × i O - 4 G N ~ --3.091 × 1 0 - 3 G N + 1 .394G 3 -- 2 . 8 2 1 G 2 + 1 . 7 9 6 G - 0 .0406

Delay (sec) Ostd : ( - 5 . 0 0 5 × i O - ~ G ~ N + 2 . 8 1 0 × 210 6 G N a --3.117 X 1 0 - 4 G N 2 + 1 . 6 8 2 x 10 - 2 G N + 4 . 8 7 0 G 3 -- 10 ,187G z + 6 . 8 6 6 G - 0 .499) 2°

DR7, : - -8.048 × I O - ~ G N + 1.080 x 10 a N q-3.002 × 10 - 4

Missed /~\~ta = s in~( 27-574Ga -- 5 8 . 0 5 5 G z Deadline + 3 9 . 6 7 0 G + 1,1.80 × 1 0 - 6 N a

Ratio - 2 . 6 2 5 × 1 0 - a N - 7.186) FRT = - - 0 . S 0 7 G ~ + 1 .993( / -- 0.521

Collision Cs ta = - -0 .437G 2 + 0 . 6 2 8 G - :l.015 × 1 0 - ' ~ N z Ratio + 0 . 0 1 2 6 N - 0.184

CRT = 5.445 × 1 0 - ' N a -- 6.258 x 1 0 - ~ N ~ + 2 . 3 7 8 × 1 0 - a N + 9,898 × 10 - a

0.92

0.91

0.99

0.99

0.99

0.99

0.98

0.91

Table 7: Avionics Traffic Model Regression Models Response

Variable

Throughput

Delay (sec)

Missed Deadline

Ratio

Collision Ratio

Regression Model

Ss td = --5.072 X l O - a G a N -- 2 . 0 8 8 G a + 2 . 6 5 8 G 2 + 0.11.63

S n T = - 1 . 4 9 6 G ~ + 2 . 1 6 5 G z + 0 .143 0.99

lgst d : ( - 0 . 4 4 4 G a N + O,763GZN 0.97 - 0 . 3 9 3 G N + 7 . 1 9 2 G 3 - 1 1 . 1 8 4 G 2

+ 5 . 4 9 7 G + 0 . 0 6 1 N - 0 .043) 2° DRT : 1 .239G a - 1 . 6 9 9 G z + 0 . 7 7 8 G 0.99

+ 4 . 4 5 9 x 1 0 - 3 E - 0.103

/;'6't d : s in2( - 0 . 3 9 8 G a N + 0 . 5 1 1 G U N 0.95 - - 0 A 3 6 G N + 5 . 7 2 6 G a -- 4 . 6 6 8 G + 1 . 4 1 5 )

/7'~T = 8 . 6 t 6 X [ ( ) -aGUN + 1 .810G a 0.98 - 2 . 6 6 6 G 2 + 1 . 2 3 5 G - 0 .180

Cstd : sinU( - 0 - 3 7 6 G a N + 0 . 6 6 1 G 2 N 0.97 - 0 . 3 4 1 G N + 0 . 3 5 7 G a -F 0 . 0 5 3 5 N + 0 . 0 3 3 6 )

C n T = 5,995 × 1 0 - z G a - 2.626 × iO -ZG z 0.97 - 5 . 1 5 0 × 1 0 - 8 N 3 + 1 . 5 2 7 × 1 0 - 4 N + 2.013 × 1.0 - 4

Model R z

0.99

regression model. In general, the larger the R 2 value, the better the model fits the observed data (l.0 maximum). All factors in the regression models and their interactions were shown to be significant to probability levels less than 0.0001 and, therefore, should be retained. One of the assumptions used in a regression is that the errors (residuals) are normally distributed [17]. Residuals are the difference between the regression model pre- diction and the simulation data. To verify normality, when the residuals are plotted in a normal quantile-quantile graph, the resulting line should be linear. The normal quantile-quantile graphs for the regression models residuals in Tables 6 and 7 were generally quite linear indicating that, in fact, the residuals were normally distributed.

Tables 6 and 7 contain the regression models for the telemetry traffic model and the avionics traffic model, respectively. These models are valid for 5 _< N < 50, 0.3 _< G _< 0.9 and E = 0, 1 where N is the number of stations in the network, G is the total offered load, and E = 0 means an ideal channel and E = 1 means an errored channel as described in Section IV.

In Section V.A we stated that RT-MAC stabilized the behavior of the response variables. To support this claim we note that RT-MAC, in general, requires fewer terms and

25 Mobile Computing and Communications Review, Volume 3, Number 2

those terms have fewer GZN v interactions than the corre- sponding IEEE 802.11 models. Specifically, we observe that F n r (telemetry traffic model) and SR:c, DnT' (avionics traffic model) are virtually independent of N. in contrast to IEEE 802.11, RT-MAC provides both better performance and a graceful degradation of performance in high network demand situations.

The simulation factor E is conspicuous in the regression models by its absence. Indeed, it only appears in the DRT. model for avionics traffic. It was found that, while usually statistically significant, the effect of an errored channel was masked by the effects of either G or JV or both. We attribute this to two reasons. First, the mean amount of time in which errors can occur is quite small compared to the error-free time. Second, especially in the case of the telemetry traffic model, the amount of data that must be retransmitted when an error does occur is relatively small and much of the time in the errored state is spent waiting for an acknowledgement. By the time the next transmission occurs, much of the time in the errored state has past. However, it should not be concluded that bit errors do not have a discernible effect on network performance--the simulation data indicates they do. Rather, we conclude that for the error model employed, the regression models were influenced to a higher degree by N and G.

VI. C o n c l u s i o n s a n d F u t u r e Work

We have presented a simple, elegant real-time enhancement to the IEEE 802.11 medium access control (MAC) protocol, RT- MAC. We have shown that by implementing a transmission control policy and sharing station BVs that dramatic performance improvements can be realized. Further, we demonstrate that network behavior in terms of mean delay, missed deadline ratio, and collision ratio is stabilized. By using RT-MAC, packet missed deadlines and mean delay became virtually independent of the number of stations in the network. Addi- tionally, we developed regression models that can be used to predict key performance measures of both IEEE 802.11 and RT-MAC networks.

We are currently investigating the performance of RT-MAC in an ad hoc network with stations that are transmitting real- time voice traffic along with an increasing load of non-real- time traffic. Further, we are studying the effect on network per- tormance of adding IEEE 802.11 stations to a network consist- ing of RT-MAC stations. We also intend to determine whether mean delay can be further improved by optimizing the CW expansion factor.

RT-MAC was evaluated using a wireless LAN. The improvements offered by our protocol, however, are not limited to wireless LANs or even to real-time data traffic. Our results extend to any time-slotted LAN, wired or wireless. Fur- ther, our enhanced collision avoidance algorithm can be implemented independent of the data being transported.

R e f e r e n c e s

[1] N. Abramson. The ALOHA system-another alternative for computer communications. In AFIPS Conference Proceedings 1970 Fall Joint Computer Conference, volume 37, pages 281-285, 1970.

[2] N. Abramson. The throughput of packet broadcast- ing channels. IEEE Transactions on Communications, COM-25(1): 117-128, January 1977.

[3] ASC/ENA Wright-Patterson AFB, OH, Wright- Patterson AFB, OH. Digital Time Division Co,z- mand/Response Multiplex Data Bus, MIL-STD- 1553B edition, September 1978.

[4] R. O. Baldwin, N. J. Davis IV, and S. F. Midkiff. Imple- mentation of an IEEE 802.1 t wireless LAN model using OPNET. In Proceedings ofOPNETvVORK'98. MIL3, Inc., June 1998.

[51] R Bhagwat, R Bhattacharya, A. Krishna, and S. Tripathi. Enhancing throughput over wireless LANs using channel state dependent packet scheduling, in Proceedings 1EEE INFOCOM '96, pages 1133-1140. Institute of Electrical and Electronics Engineers, 1996.

[6] R Bhagwat, R Bhattacharya, A. Krishna, and S. Tripathi. Using channel state dependent packet scheduling to improve TCP throughput over wireless LANs. Wireless Networks, 3:91-102, 1997.

[7] V. Bharghavan. Performance evaluation of algorithms for wireless medium access. In IEEE International Computer Pe@)rmance and Dependability Symposium. IPDS'98, pages 86-95. Institute of Electrical and Elec- tronics Engineers, 1998.

[8] G. Bianchi, L. Fratta, and M. Oliveri. Performance evaluation and enhancement of the CSMA/CA MAC protocol for 802.11 wireless LANs. The Seventh tEEE bzterna- tional Symposium on Personal, Indoor and Mobile Radio Communications PIMRC '96, pages 392-396, October 1996.

[9] C. R. Braun. Wireless LAN standard proposal highlights the need for speed. Wireless Systems Design, 3(9): 13-16, September 1998.

[10] E Call, M. Conti, and E. Gregori. IEEE 802.t t wireless LAN: Capacity analysis and protocol enhancement. In INFOCOM '98, Conference on Computer Communi- cations, pages 142-149. Institute of Electrical and Elec- tronics Engineers, 1998.

[11] T. Carpenter, K. Driscoll, KfHoyme, and J. Carciofini. ARINC 659 scheduling: Problem definition. In 1994 Real-Time Systems Symposium, pages 165-169. Institute of Electrical and Electronics Engineers, 1994.

[t2] B. Crow, I. Widjaja, J. G. Kim, and R Sakai. Perfor- mance of IEEE 802.11 wireless local area networks. In Proceedings of the SPIE, volume 2917, pages 480-491. The International Society of Optical Engineers, 1996.

[13] R. Dube, C. Rais, and S. K. Tripathi. Improving NFS performance over wireless links. IEEE Transactions on Computers', 46(3):290-298, March 1997.

[14] Editors of IEEE 802.11. Wireless LAN Medium Access Control (MAC) andPhysical Layer (PHY) Specifications,

26 Mobile Computing and Communications Review, Volume 3, Number 2

Draft Sta~da~d 802.11, P802. ] ]/D6.1. Institute of Elec- trical and Electronics Engineers, Inc., New York, May 1997.

115] J. Etlsberger, D. Hogrefe, and A. Sm'ma. SDL, Pormal Object-Oriented Language for Communicating Systems. Prentice Hall Europe, Hertl~rdshire, UK, 1997.

[16] L. Kleinrock F. S. Tobagi. Packet switching in radio channels: Part III-polling and (dynamic) split-channel reservation multiple access. IEEE Transactions on Com- munications, COM-24(8):832-845, August 1976.

[ 17] R. Jain. The Art of Computer Systems Pelformance Anal- ysis. John Wiley and Sons, Inc., New York, 1991.

[18] L. Kleinrock. Queueing Systems Volume 1: 77~eory. John Wiley and Sons Inc., New York, 1975.

[19] L. Kleinrock. Queueing Systems Volume 2. Computer Applications. John Wiley and Sons, Inc., New York, t976.

[20] L. Kleinrock. Principles and lessons in packet comnm- nications. Proceedings of the IEEE, 66(11): 1320--1329, November 1978.

[21] L. Kleinrock and M. O. Schotl. Packet switching in radio channels: New conflict-free multiple access schemes. IEEE Transactions on Communications, COM- 28(7):1015-1029, July 1980.

[22] L. Kleinrock and E A. Tobagi. Packet switching in radio channels: Part I: CSMA modes and their throughput- delay characteristics. IEEE Transactions on Communi- cations, COM-23(12): 1400-1416, December 1975.

[23] J. E Kurose, M. Schwartz, and Y. Yemini. Multiple- access protocols and time-constrained communication. Computing Surveys, 16(1):43-70, March 1984.

[24] T. Liu, J. A. Silvester, and A. Polydoros. Performance evaluation of R-ALOHA in distributed packet radio networks with hard real-time communications. In 1995 IEEE 45th Vehicular Technology Co@rence, volume 25, pages 554-558. Institute of Electrical and Electronics Engineers, Inc., July 1995.

[25] M.H. MacDougall. Simulating CompuWr Systems: Tech- niques and Tools. The MIT Press, Cambridge, MA, 1992.

[26] N. Malcohn. Hard Real-l~me Communications in High Speed Networks. PhD thesis, Texas A&M University, December 1994.

[27] N. Malcolm and W. Zhao. Hard real-time communications in multiple-access networks. Real-7]'me Systems, 8(1):35-77, January 1995.

[28] MIL 3, Inc., 3400 International Drive, NW Washington D.C., 20008. OPNETModeler, 1997.

[29] J. L. Sobrinho and A. S. Krishnakumar. Real-time traffic over the IEEE 802.11 medium access control layer. Bell Labs Technical Journal, pages 172-187, 1996.


[30] The SAS System. SAS Institute, Inc. Cary, NC.

[31] F. S. Tobagi and L. Kleinrock. Packet switching in radio channels: Part II-the hidden terminal problem in carrier sense multiple-access and the busy-tone solution. IEEE Transactions on Communications, COM-23(12):1417- 1433, December 1975.

[32] F. S. Tobagi and L. Kleinrock. Packet switching in radio channels: Part IV-stability considerations and dynamic control in carrier sense multiple access. IEEE Transac- tions on Communications, COM-25(10): 1103-1119, Oc- tober 1977.

[33] E S. Tobagi and L. Kleinrock. The effect of acknowl- edgment traffic on the capacity of packet-switched radio channels. 1EEE 7?ansactions on Communications, COM-26(6):815-826, June 1978.

[34] M. A. Visser and M. E1 Zarki. Voice and data transmission over an 802. l 1 wireless network. In 6th IEEE In- ternational Symposium on Personal, bldoo~ and Mobile Radio Communications, volume 2, pages 648-652. Insti- tute of Electrical and Electronics Engineers, 1995.

[35] J. Weinmiller, H. Woesner, and A. Wolisz. Analyzing and improving the IEEE 802.1 l-MAC protocol for wireless LANs. In MASCOTS 96. Proceedings of the Fourth Inter- national Workshop on Modeling, Analysis, and Simula- tion of Computer and Telecommunication Syswms, pages 200-206, February 1996.

[36] W. Zhao and K. Ramamritham. Virtual time CSMA protocols for hard real-time communication. IEEE Transac- tions on Software Engineering, SE-13(8):938-952, Au- gust 1987.

Biographies

Rusty O. Baldwin is a captain in the U.S. Air Force and a Ph.D. student in electrical engineering at Virginia Tech. He received his B.S.E.E. degree (with honors) from New Mexico State University in 1987, and his M.S. degree in Computer Engineering from the Air Force Institute of Technology in 1992. His research interests include computer communication protocols, software engineering, and wireless networking.

Nathaniel J. Davis IV is an Associate Professor in the Bradley Department of Electrical and Computer Engineering at Virginia ]~ch. He received his B.S. and M.S. degrees from Virginia Tech and his Ph.D. degree from Purdue University. His research interests include wireless communications networks, computer performance evaluation, and high-performance computer architectures.

Scott E Midkiff is an Associate Professor in the Bradley Depamnent of Electrical and Computer Engineering at Virginia Tech. He received his B.S.E. degree in 1979 from Duke University, his M.S. degree from Stanford University in 1980, and his Ph.D. from Duke University in 1985. His research interests include wireless networking, network man- agement related to quality-of-service and capacity planning, performance evaluation, and multimedia applications.

27

Documents

A Real-time Medium Access Control Protocol for Ad hoc Wireless