1: \section{EDCA Parameter Analysis}\label{sec:analysis}
2:
3: The main motivation behind the design of EDCA is providing a
4: framework where the medium access of coexisting flows can be
5: prioritized according to their application layer traffic class.
6: The main intention is to provision QoS for real-time flows by
7: prioritizing their access over best-effort and background traffic.
8: On the other hand, as also shown in the literature,
9: uplink/downlink unfairness problem can be combatted by the
10: assignment of AC-specific EDCA parameters with respect to the
11: traffic direction instead of the traffic class. In this section,
12: we point out the fact that special care must be taken in the
13: design of such framework so that the QoS requirements of the
14: coexisting real-time flows can be maintained (the main intention
15: of QoS provisioning behind the EDCA design is not
16: jeopardized)\footnotemark{}.
17:
18: \footnotetext{The 802.11e standard suggests the use of an
19: admission control algorithm for QoS provisioning. In an ideal
20: scenario, the admission control algorithm prevents the access of a
21: real-time station if its admittance to the network can degrade the
22: overall QoS. This also means that, in an ideal scenario, QoS is
23: preserved, all real-time stations are nonsaturated, and unfairness
24: problem does not exist for QoS stations. Therefore, in this paper,
25: we consider the best-effort traffic that no admission control is
26: or can be applied.}
27:
28: The solution for resolving the uplink/downlink unfairness problem
29: using EDCA parameter differentiation is pretty clear: Prioritize
30: the access of the given AC at the AP. This can simply be achieved
31: by assigning the specific AC at the AP \textit{i)} a lower AIFS
32: value, \textit{ii)} a lower CW, \textit{iii)} a higher TXOP limit,
33: or \textit{iv)} any joint combination of these, when compared to
34: the assigned parameters of the specific uplink AC. The challenge
35: is to find the parameters that would provide weighted channel
36: access while preserving QoS demands of higher priority realtime
37: flows.
38:
39: Let's first briefly review the effects of EDCA parameter selection
40: on the achievable uplink/downlink throughput ratio within an AC.
41:
42: \begin{itemize}
43: \item Each AC can either transmit or start decrementing its backoff
44: counter if the channel is detected to be idle for the duration of
45: the AC-specific AIFS value \cite{802.11e}. This means that the
46: access for ACs with higher AIFS values are further delayed
47: compared to the ACs with lower AIFS values every time the channel
48: becomes busy. At low channel load, the effect of AIFS on
49: prioritization is not significant, since the backoff countdown is
50: not frequently interrupted by other transmissions. Conversely, at
51: high channel load, AIFS prioritization becomes a significant
52: factor. Every time the channel is grabbed, this directly means a
53: further delay on the stations with lower priority (i.e., larger
54: AIFS) when compared to the stations with higher priority.
55:
56: \item The stations pick a backoff value uniformly distributed between
57: 0 and the current CW size and complete a backoff countdown before
58: transmission \cite{802.11e}. Upon gaining access to the medium,
59: each AC may carry out multiple frame exchange sequences as long as
60: the total access duration does not go over its TXOP limit
61: \cite{802.11e}. The channel access ratio between uplink and
62: downlink within an AC varies almost linearly with respect to the
63: selection of $CW_{min}$\footnotemark{}\footnotetext{As
64: \cite{802.11e} defines, the initial value of AC-specific CW is
65: $CW_{min}$. At every retransmission the CW is doubled, up to
66: $CW_{max}$.} and $TXOP$ values in asymptotical conditions
67: (saturation). Following our analytical calculation in
68: \cite{Keceli08_ICCEDCA}, the approximate channel access ratio
69: $U_{i,u/i,d}$ between uplink and downlink within AC$_{i}$, namely
70: AC$_{i,u}$ and AC$_{i,d}$, can be calculated as
71: \begin{equation}
72: \label{eq:U_{i/j}}
73: U_{i,u/i,d}\cong\frac{CW_{min,i,d}N_{TXOP_{i,u}}}{CW_{min,i,u}N_{TXOP_{i,d}}}
74: \end{equation}
75:
76: \noindent when $AIFS_{i,u}=AIFS_{i,d}$ and both directions are
77: saturated. Note that $N_{TXOP_{i}}$ denotes the maximum number of
78: packets that AC$_{i}$ can fit into one TXOP.
79: \end{itemize}
80:
81: Simply by using (\ref{eq:U_{i/j}}), we can calculate a set of
82: $CW_{min}$ and $TXOP$ values for AC$_{i,d}$ that would
83: approximately achieve a predetermined throughput ratio
84: $U_{i,u/i,d}$ for given $CW_{min}$ and $TXOP$ values of
85: AC$_{i,u}$, and vice versa. On the other hand, the throughput
86: ratio achieved by AIFS differentiation is yet to be approximated
87: via a simple linear relationship as it can be done in
88: (\ref{eq:U_{i/j}}) for CW and TXOP. Therefore, in this work, we
89: consider joint CW and TXOP differentiation in provisioning
90: weighted fair access. AIFS differentiation is only used between
91: ACs to provide prioritization between realtime and best-effort
92: flows.
93:
94: We carried out a simulation-based analysis to further analyze the
95: effects of CW and TXOP differentiation on both fair access and QoS
96: provisioning. We consider a scenario with two active ACs. For both
97: ACs, we use a traffic model with Poisson packet arrivals. The
98: transport layer protocol is UDP for both ACs. We use 11 Mbps
99: 802.11b PHY and assume that the wireless channel is errorless.
100:
101: The high priority traffic uses AC$_{3}$ with EDCA parameters
102: $AIFS=2$, $CW_{min}=7$, $CW_{max}=15$, $TXOP=1.504~ms$ both at the
103: AP and the stations (as suggested in \cite{802.11e}). We consider
104: 5 uplink and 5 downlink high priority flows generated at 250 kbps.
105: We intentionally do not saturate the high priority AC, so that the
106: lower priority ACs do not starve and the effects of CW and TXOP
107: selection can be observed. This also corresponds to a practical
108: scenario since the traffic load should be well controlled and an
109: admission control algorithm should keep the high priority AC
110: nonsaturated to support parameterized QoS \cite{Zhai05,
111: Inan07_multimediacap_trep}.
112:
113: The low priority AC is considered to be serving best-effort
114: traffic. We set the traffic load so high that the low prority AC
115: is saturated both at the stations and the AP. This is also a valid
116: assumption to analyze the worst-case scenario, since no admission
117: control is applied for best-effort traffic category in practice.
118: We consider four different cases in assigning the EDCA parameters
119: of the AP and the stations for the low priority traffic.
120:
121: \begin{itemize}
122: \item Default: Both the AP and the stations use the same parameters
123: which are tentatively set as $AIFS=3$, $CW_{min}=31$,
124: $CW_{max}=511$, $TXOP=0$.
125: \item TXOP differentiation: The AP is assigned a TXOP regarding
126: the total number of downlink flows $n_{d}$, i.e., $TXOP=n_{d}\cdot
127: T_{exc}$, where $T_{exc}$ is the time required to complete a data
128: frame exchange (including MAC/PHY overhead). Note that this is
129: similar to the approach proposed in \cite{Leith05},
130: \cite{Freitag06}.
131: %\item CW differentiation: The stations are assigned a $CW_{min}$
132: %regarding the total number of downlink flows $n_{d}$, i.e.,
133: %$CW_{min} = 31\cdot n_{d}$.
134: \item CW differentiation: The AP is assigned a smaller $CW_{min}=7$.
135: The stations are assigned a $CW_{min}$ regarding the total number
136: of downlink flows $CW_{min} = 7\cdot n_{d}$.
137: %This is a worst-case
138: %scenario since $CW_{min}$ values of low priority traffic at the AP
139: %is equal to what high priority AC uses.
140: \item Joint CW and TXOP adaptation: We employ our joint adaptation
141: approach, WFA, as proposed in this paper in the sequel.
142: \end{itemize}
143:
144: Note that in the last three cases, the parameters are set so that
145: a utilization ration of 1 between uplink and downlink can be
146: approximately achieved.
147:
148: Fig. \ref{fig:jainfairness_qos_11udp} and Fig.
149: \ref{fig:totalthroughput_qos_11udp} show the fairness index and the
150: total throughput for the best-effort AC, respectively.
151: %Since the QoS stations achieve their demanded bandwidth,
152: %we do not present these results separately.
153: Fig. \ref{fig:delay_qos_11udp} and Fig. \ref{fig:jitter_qos_11udp} show the average delay and the average jitter experienced
154: by QoS flows for increasing number of low priority uplink and
155: downlink stations, respectively.
156: %Due to space limitations, we present the jitter
157: %results for QoS flows in a technical report
158: %\cite{Keceli08_fairEDCA_trep}.
159: We can extract the following
160: insights from the presented simulation-based analysis.
161:
162: \begin{itemize}
163: \item As shown in Fig.
164: \ref{fig:jainfairness_qos_11udp}, static CW and TXOP
165: differentiation cannot maintain fair access as a result of the
166: fact that the channel access ratio as calculated by
167: (\ref{eq:U_{i/j}}) is only an approximation. The design of an
168: analytical model which captures all network details in order to
169: calculate the exact parameters is hard and complex. Dynamic
170: adaptation as proposed in this paper simply preserves weighted
171: fair access.
172: \item As shown in
173: Fig. \ref{fig:totalthroughput_qos_11udp}, when compared to the
174: default case, both TXOP differentiation and CW differentiation
175: improve channel utilization. In the former case, channel
176: contention overhead is decreased by the use of TXOP. In the latter
177: case, the stations are assigned larger $CW_{min}$ values so that
178: the collision overhead is decreased while the downlink enjoys a
179: higher channel access rate with the assigned smaller $CW_{min}$
180: value.
181: %The proposed joint CW and TXOP adaptation scheme can
182: %maintain high channel utilization while providing fair access.
183: \item As shown in Fig. \ref{fig:delay_qos_11udp}, as the number
184: of best-effort flows increases, employing TXOP differentiation at
185: the AP for low priority traffic jeopardizes the QoS of high
186: priority flows (the average delay increases exponentially). If a
187: packet belonging to a QoS flow arrives while the channel is busy
188: because of a best-effort transmission, the QoS packet has to wait
189: a long time until the transmission is completed. On the other
190: hand, in the case of CW differentiation, when best-effort flows
191: access the channel, they hold the channel for a much shorter
192: duration at every access which means a smaller access overhead for
193: the QoS stations.
194: \item A smaller CW selection at the AP for low priority flows
195: does not degrade QoS of higher priority flows in the same order of
196: TXOP differentiation. In the specific scenario, the downlink
197: best-effort flows use the same $CW_{min}$ as the QoS flows are
198: assigned. The differentiation is still maintained via different
199: AIFS values. Moreover, the access frequency for the stations are
200: decreased since they use larger CW values (compared to the default
201: EDCA and TXOP differentiation cases). Conversely, the total
202: throughput for the best-effort traffic increases (due to lower
203: collision overhead) and the QoS stations experience a low packet
204: delay.
205: \item A similar discussion as for delay holds on the jitter of QoS flows as per the results
206: presented in Fig. \ref{fig:jitter_qos_11udp}.
207: \end{itemize}
208:
209: Fig. \ref{fig:jainfairness_qos_11tcp}-\ref{fig:jitter_qos_11tcp}
210: show the results when best-effort flows employ TCP. As can be seen
211: from Fig.
212: \ref{fig:totalthroughput_qos_11tcp}-\ref{fig:jitter_qos_11tcp},
213: similar discussions on the comparison hold for throughput of the
214: best effort flows, delay and jitter experienced by QoS flows.
215: However, as shown in Fig. \ref{fig:jainfairness_qos_11tcp}, both
216: TXOP differentiation and CW differentiation provide fair access
217: among TCP flows (different than UDP). These schemes implicitly
218: make use of the results of capture effect in fair access
219: provisioning. As a result of capture effect, the EDCA parameters
220: settings favors the downlink access in both CW differentiation and
221: TXOP differentiation. As the results present, while this causes
222: unfair access between uplink and downlink UDP flows, fair medium
223: access is still maintained among TCP flows (when TCP does not
224: employ the delayed ACK mechanism). The reasoning behind this
225: behavior is actually what motivates the design of proposed EPDA
226: algorithm in the sequel and will be described in detail in Section
227: \ref{subsec:EPDA_TCP}.
228: %(Note that although both CW differentiation
229: %and TXOP differentiation provide a fair MAC layer for download and
230: %upload by giving either n times more TXOP to the AP or n times
231: %more CW to the STAs , the AP gets higher transmit opportunities
232: %mostly due to the capture effect.)
233:
234: Similar discussions hold when the stations use 54 Mbps 802.11g PHY
235: layer as presented in Fig.
236: \ref{fig:jainfairness_qos_54udp}-\ref{fig:jitter_qos_54tcp}.
237:
238: This analysis motivates the joint use of CW and TXOP
239: differentiation for efficient and fair medium access. A multiple
240: packet exchange in a TXOP improves channel utilization by
241: decreasing contention overhead. On the other hand, as our analysis
242: implies, the TXOP limit assigned should not go over a threshold
243: for concurrent fair access and QoS provisioning. Although we
244: provide the results for the proposed WFA scheme in Fig.
245: \ref{fig:jainfairness_qos_11udp}- \ref{fig:delay_qos_11udp}, we
246: provide the discussion on WFA performance in Section
247: \ref{sec:simulations} after the framework is described in Section
248: \ref{sec:framework}.
249:
250:
251: %The main building block of the proposed framework for fair access
252: %provisioning is the EDCA parameter adaptation block. As it will be
253: %described in more detail in Section \ref{sec:framework}, the
254: %proposed parameter adaptation mainly prioritizes the AP access on
255: %the best-effort ACs by assigning smaller CW and larger TXOP values
256: %to the specific AC. In this section, we evaluate the effects of
257: %such prioritization employed for the best-effort traffic at the AP
258: %on the performance of coexisting real-time traffic in the network
259: %in terms of average delay and jitter.
260:
261: %In the limiting scenario, assume that the best-effort AC of the AP
262: %has the same AC-specific parameter set of an AC that serves
263: %real-time flows. This also corresponds to the case that the total
264: %number of ACs serving real-time flows is just increased by one.
265:
266: %Analytical model discussion if decided to be included, mainly
267: %refer to the previous work
268:
269: %access delay and per saturation analysis, showing the effects of
270: %the addition of one station with the same EDCA parameters
271:
272: %wrap up with CW and TXOP (and maybe AIFS) results
273: