Token Enabled Multiple Access (TEMA) for Packet Transmission in High Bit Rate Wireless Local Area Networks Saeid Akhavan Taheri
Anna Scaglione
Elecrical & Computer Eng. Dept., University of New Mexico Email:
[email protected]
School of Elecrical & Computer Eng., Cornell University Email:
[email protected]
Abstract—In this paper we illustrate a novel Wireless Local Area Network Multiple Access Control (WLAN MAC) protocol whose allocation policy adopts a strategy similar to the token ring: the mobile stations access the channel upon receiving the token and release the channel by passing the token to the next user. The traffic is divided in two classes: Guaranteed Bandwidth (GB) and Best Effort (BE). The base station indirectly controls the traffic access by allotting the tokens and the bandwidth is shared by the same users with a decentralized contention-less protocol in case of the GB users and on a contention basis for the BE users. The MAC protocol proposed the physical layer adopts a Multicarrier Code Division Multiple Access multiplexing technique, which mitigates the linear distortions of broadband transmission and allows to divide the frequency and time resources in elementary units that can be associated to the ’token’. The specification on the average and peak delay and rate are guaranteed by limiting the number of users passing the token and fixing a deadline to release the token. At the same time, similar to a reservation based protocol, the token strategy avoids collisions and is able to guarantee efficiently the integrity of the transmitted data. Index Terms—Wireless Networks, Multiple Access
I. I NTRODUCTION N case of bursty data sources, multiple access schemes that perform a static channel assignment, such as TDMA, FDMA, CDMA, SDMA, are known to be inefficient compared to contention based multiple access protocols. ALOHA-type random access protocols have been proposed to solve the problem of “contention” based systems [1]. However, in high traffic conditions ALOHA-type (slotted and un-slotted) random access protocols yield relatively low throughput, due to the high collision probability. The carrier Sense Multiple Access Protocols (CSMA with collision detection, non-persistent, 1persistent and p-persistent) offer a more efficient alternative to ALOHA protocols. In CSMA the station listens to the carrier and when the channel is idle, transmits a packet [11]. In particular, CSMA with collision detection (CD), which is used as a standard Ethernet protocol, minimizes the channel downtime. CSMA/CD yields a high throughput when the propagation delays (τd ) are relatively small in comparison to the packet duration (2τd < Tp ) [1]. An alternative scheme is the so called Demand Assigned Multiple Access (DAMA) where a separate request channel is used by the users to request a specific rate. The request is processed by a central master station or a common algorithm running in each terminal [1]. DAMA is a dynamic reservation-based algorithm and it is effective when the set of active transmitters changes rapidly. Scheduling at packet level according to the demanded QoS have been considered in several works: a comprehensive survey on packet scheduling in wireless networks is provided in [3].
I
In this paper we propose a distributed multiple access control scheme, named Token Assignment Multiple Access (TEMA), designed to handle efficiently heterogenous classes of traffic in heavy traffic load. Our scheme mainly is based on the polling and the token ring concept. Polling multiple access protocols have been considered for local communication networks for wired media [13], [18], [9]. In these protocols the terminals are interrogated in a cyclic order, by either a central controller or, in a distributed fashion, by using specialized polling query/response (token) packets [14]. The stations are organized into a ring (they are, in fact, named token-ring networks and physically they can be organized in a tree shaped cable) and each station knows the addresses of the stations at its left and right. When the logical ring is initialized, the highest numbered station sends the first frame. Subsequently, the station passes the permission to its immediate neighbor by sending to the neighbor a special control frame called a token. Token ring are one of the most reliable LAN in the industry [7]. However, wired token ring networks have the disadvantage that a broken section in the ring cable causes an outage of the whole network. Also, in comparison to the other MAC protocols, the token-passing MAC has control overhead [10]. To analyze the TEMA network performance we extend the priority-based token ring to the wireless setting and borrow results from [17], [18], which derive the performance of priority polling schemes designed to provide isochronous and asynchronous type services for wired media. For the QoS specifications we provide a coarser separation by identifying only two classes of service, Guaranteed Bandwidth (GB) and Best Effort (BE) classes. TEMA is initialized is with a random access protocol: According to the demanded QoS the base station initializes the rings, by allotting a subchannel (token) to a set of users (ring). The demanded contention free part for high QoS users is constructed in a distributed manner, by passing the token from user to user. Using the rest of the resources through contention based scheme we do not waste bandwidth. The rest of this paper is organized as follows. Section I is dedicated to the description of the basic TEMA idea and is followed by the description of the specific MC-CDMA physical layer that we choose to support the TEMA architecture. Section IV shows the performance analysis and comparison between existing multiple access modes. II. TEMA: T OKEN E NABLED M ULTIPLE ACCESS Consider N users Ui ordered by i, i = 1, . . . , N wishing to communicate with a base station through a multipleaccess communication channel. Our idea is to logically organize groups of users in subnetworks (interconnected wireless
0-7803-7400-2/02/$17.00 © 2002 IEEE
1913
token rings), so that with the polling discipline which is not centralized each user polls the consecutive user in the carousel (see figure 1). Each of the N users in the ring are queried by an algorithm in a distributed fashion. If the answer to this query is positive, the channel is completely assigned to the user for data transmission. Upon completion the transmission or giving negative answer to the query, the control polls the next user in order. A station is served until either 1) its buffer is emptied, or 2) a specified time slot to the user is over, whichever occurs first. Each token corresponds to a physical MC-CDMA subchannel which is guaranteed to have a certain average rate and satisfy a probability of error bound, as described in detail in Section III.
one rotation of the high-priority token in the time left the low-priority token will rotate. C. Non-homogenous traffic, high quality of service and bursty traffic. In this scheme the token passing overhead is avoided for the bursty data by alternating the users that pass the token with a contention based period where all bursty data sources contend the bandwidth with a CSMA strategy. Figure 2 shows the activity of the channel for these three (K) denote the service time for ith user in kth schemes. Let Ti (K) reception of token, Ci be the length of the kth token cycle (K) for user i and Wi be the so called walk time, which includes (K) the token passing overhead. Similar to [13], we define Qi to be transmission quota, i.e., for each token cycle the ith user (K) (K) message duration is limited, Ti ≤ Qi . P (K) N c = From fig. 2, denoting by W we can derive i Wi (K) (K) for each of the the following relations between Ci and Ti proposed schemes: For scheme 1 shown in Fig. 2 A),
"! $#&%'
(K)
c+ =W
Ci
M X
(K)
Tj
+
i X
j=i
Fig. 1. Polling with cyclic visit ordering for N users and one server.
(K+1)
Tj
.
(1)
j=1
For scheme 2 shown in Fig. 2 B),
To simplify the scheme’s analysis and having analogy with the existing schemes for fiber distributed data interface (FDDI) [5], only two different classes of QoS are considered: the guaranteed bandwidth (GB) class, for which the requirement is either that P r(delay > D) < ² or P r(end − to − end packet loss) < ² for a given ², and the best effort (BE) class, for which there is no requirement. Several priority-based token schemes for wired networks have been studies [17]. In particular, one of our schemes (the second one presented below) shares similarities with the delayconstrained scheme in [17], where messages are divided in high priority and low priority messages. By defining two types of tokens, high level and low level, the system allows all users in the ring to send their high level messages at first and then low level messages. This scheme is not well tailored for wireless communications, since the type of load and bandwidth usage and probability of error are different. We propose three different token-schemes applicable to the wireless media for: A. Homogenous traffic; in this scheme all users demanding a single type of service with high quality (e.g., broadband multimedia). Hence, only one type of token (priority) is considered. To satisfy the tight delay requirements, a limited dwell time policy is applied to this scheme. B. Non-homogenous traffic; in this scheme there are non-bursty to moderately-bursty users which require higher and lower QoS respectively. This scheme is applied when the users can have two different type of demanded QoS. In the ring first all high-priority users are served while the remaining resources are dedicated to the low priority packets. In this scheme, highpriority and low-priority tokens are employed. After
(K) Ci
c+ =W
NX −M
(K) Tj
+
M X
j=i
(K+1) T2j
+
j=1
i X
(K+1)
Tj
. (2)
j=1
Finally, for scheme 3 shown in Fig. 2 C), (K)
Ci
NX −M
c+ =W
(K)
Tj
F X
+
j=i
(CSM A)
Tj
+
j=i
i X
(K+1)
Tj
. (3)
j=1
Each token is not only associated to a physical set of subcar(K)
T1
(K)
W
(K)
(K)
T2
1
1
TN-M
N-M
2
C
(K+1)
TM
M
1 QN-M-s+1
QN-M-s
Q1
(K+1)
T1
QN-M-s+2
QN
(K) 1
C
(K) 2
A) GB traffic. (K)
T11
(K)
W
(K)
(K)
T12
11
T1N-M
12
11
(K+1)
T21
1N-M 21 QN-M-s
Q1 C
QN-M-s+1QN-M-s+2
(K+1)
T2M
2M QN
(K) 1
B) GB traffic with two differenent priorities
(K)
T1
(K)
W
(K)
(K)
T2
1
1
TN-M
N-M
2
T
----Random Access---
QN-M-s
Q1 C
QN
(K) 1
C) GB and BE traffic
Fig. 2. Timing diagram for different token passing approaches; Guaranteed Bandwidth (GB) and Best Effort (BE) traffic.
1914
B/M
B
(K)
riers but also to a fixed frame duration, the token cycle Ci . The maximum number of users supported in a ring and the to(K) ken cycle Ci depend on the QoS that the ring has to support and on the bandwidth limitations of the physical layer. With reference to Fig. 2A, to guarantee P r(delay < D) < ², we assign N users to be the primary members of the ring, but indeed only N −M of them are using resources. M is a random variable indicating the number of users are silent in the ring. The primary ring members are guaranteed to receive the token (K) (K) (K) after at most the time equal to Ci +Qi +Wi . In scheme 3, the M vacancies goes for the M other members, that have no QoS guarantees, share the remaining time using random access. They can experience any delay or packet loss, but this is compliant with the type of service that they require. In practice, all the time which is left over by the primary users, is filled with the BE users data periodically on a contention basis. In fact, even if the number of ring primary users reaches the maximum number for which it can be guaranteed that P r(delay < D) < ²) the (K) GB users will not frequently use their entire quota Qi , unless they are Constant Bit Rate (CBR) sources. As a result, the fraction of total time of the token cycle in which the GB users are idle, i.e.: X (K) ∆(K) := C (K) − Ti (4)
Ts
PTs Fig. 3. Partition of the resources in the MC-CDMA system.
MC-CDMA strategy we distribute the spread data over the entire grid in Fig. 3. If P is the number of slots in one frame there are P M subcarriers per frame. To eliminate possible interference between different rings and thus the need of power control, we partition the entire set of subcarriers T in Fig. 3 mapping each token onto a distinct subset Si , i.e. such that: \ [ Si j T and Si ≡ {∅}. (5) i
This strategy is illustrated by Fig. 4, where different colors correspond different codes:
III. T HE MC-CDMA P HYSICAL L AYER Although TEMA can be adapted to other physical layer platforms, we will see next that MC-CDMA offers greater flexibility in the assignment of the transmission resources. Multicarrier communications seem to offer a cost effective solution for the new generation of broadband wireless communications [4]. MC-CDMA combines the good properties of the two emerging multiplexing and modulation techniques for wideband transmission, namely: CDMA, chosen for 3G cellular communications (UMTS and IMT-2000) and OFDM, which adopted in the wireless LAN standards IEEE 802.11 [11]and HIPERLAN [8]. A good comparison between the different multiplexing alternatives and the advantages of the MC-CDMA approach can be found [2]. In III-A we derive the basic MC-CDMA design parameters that need to be used to determine the physical resources associated to one token. Our approach allows to construct tokens that deliver the desired amount of bits/sec. satisfying the error rate bound, despite the frequency selective timevarying nature of the wireless medium, enhancing the robustness of the proposed strategy. A. The Token-Codes It is convenient to utilize a block transmission technique and associate to each token a set of signature codes that spread one block of Q data at a time. To exploit fully the flexibility of the
B/M
B
i
can be directly utilized by the BE users. Due to the distributed nature of the control mechanism, and because of the QoS and buffer limitations, it is impossible to handle all user’s demand within a single ring. The logical option is using a multi-ring system, where several non-conflicting rings organize the access to the resources.
i
Ts PTs
Fig. 4. Different MC-CDMA Token-Codes.
Let us denote by F i the Ki ×Ni matrix of MC-CDMA codes and by si [k] the block of information symbols transmitter at time k over the ith token. The superposition of the spread data is: xi [k] = F i si [k], (6) and the entries xi [k] are transmitted on the set Si containing Ki subcarriers. From our discussion it follows that, Ni ≤ Ki ≤ P M,
(7)
and the time necessary to transmit the Ni symbols is at most the frame duration Tf : Tf = P Ts =
P (M + L) B
(8)
where, as mentioned before, B is the total available bandwidth M is the number of subcarriers and L is the duration of the cyclic prefix. L has to be greater then the channel delay spread and maximum asynchronism between users. Fixing M to be greater than the delay spread of the channel, L = M would sacrifice efficiency but render block synchronization unnecessary. Assuming the receiver carrier is synchronized, the outputs of the FFT filters corresponding to the token-subcarriers are the entries of the following Ki × 1 vector:
1915
y i [k] = H i F i si [k] + v i [k],
(9)
where v i is additive white gaussian noise (AWGN) with variance N0 and H i is a diagonal with diagonal entries {H i }k,k = Hi (fik ) with fik ∈ Si . Assuming that the channel is known only at the receiver side and using symbols that belong to the same constellation of size Di , the token bit rate is (service rate) µi = Ni log2 (Di )/Tf bits/sec,
(10)
and if the target rate is µi ≥ µ ¯ then: Di ≥ 2µ¯ Tf /Ni .
(11)
service (transmission) average rate of the token code µ ¯ is normalized to one. The service administrator serves station i only for a limited time Qi (t) if there is any demand from station i. ki messages out of those found at the pooling instant are served continuously at station i. A station is served until either 1) the buffer is emptied, or 2) a specified number of messages or packets are served, whichever occurs first. Then, the token in finite reply interval wi goes to inspect station i + 1. We denote by x ¯i and by x ¯2i respectively the mean and the second order moment of the service time for the packets in the ithe queue. The utilization for user i ρi and the system utilization ρ are:
The Qos specification in terms of BER can be satisfied assuming that the ISI is approximately Gaussian and using the following bounds on symbol the error probability, where βi is the signal to noise ratio at the detection level: QAM
Ps (i)
