Network Utility Maximization for overcoming inefficiency in - PowerPoint PPT Presentation

Network Utility Maximization for overcoming inefficiency in multirate wireless networks Andr´ es Ferragut Jos´ e Garc´ ıa Fernando Paganini Mate Research Group Universidad ORT Uruguay RAWNET Workshop, 4th June 2010

Motivation Wireless network are nowadays prevalent for Internet Access. The IEEE 802.11 protocol is the most used wireless access mechanism. We want to understand the resource sharing this protocol produces in the network. Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 2 / 48

Motivation Wireless network are nowadays prevalent for Internet Access. The IEEE 802.11 protocol is the most used wireless access mechanism. We want to understand the resource sharing this protocol produces in the network. Some questions we want to answer: Which is the throughput obtained by TCP connections in a 802.11 environment? Which is the resource allocation when multiple transmission rates are present? Can we do better? Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 2 / 48

Our contributions We characterize some of the inefficiencies of TCP over 802.11 in a downlink scenario: Protocol overheads. The existence of concurrent multiple transmission rates. We propose a new allocation based on Network Utility Maximization (NUM). We present an AQM policy that enables the new allocation. We discuss how to generalize these algorithms to more complex scenarios. Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 3 / 48

Contents 1 The inefficiencies of TCP over 802.11. 2 TCP resource allocation in a multirate wireless environment 3 A more efficient resource allocation for a single cell 4 A word on more general topologies 5 Implementation and simulations 6 Conclusions Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 4 / 48

Contents 1 The inefficiencies of TCP over 802.11. 2 TCP resource allocation in a multirate wireless environment 3 A more efficient resource allocation for a single cell 4 A word on more general topologies 5 Implementation and simulations 6 Conclusions Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 5 / 48

Protocol overheads IEEE 802.11 uses DCF as the main access mechanism to the shared medium. This imposes some fixed and random times a station must comply with (DIFS, SIFS, Backoff slots, etc.) Moreover, transmitting a packet of size L bits involves some amount of overhead due to MAC and PHY headers (i.e. PLCP). In a downlink scenario these overheads predominate over the collisions. Therefore, the real transmission rate of packets is not “54Mbps” but less. How much? Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 6 / 48

MAC level rates The works pioneered by Bianchi [2] and followed by Kumar et. al. [5] established a formula for the MAC level rates. In the downlink case (where the station that predominantly accesses the medium is the AP) it takes the form: i = L L C 0 = CW min L T i σ + T 0 i + P HY i 2 L is the packet size, CW min is the minimum contention window. σ is the slot time. T 0 i accounts for the fixed time overheads in transmission. PHY i the modulation rate. In the following, we assume L = 1500 bytes. Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 7 / 48

Typical 802.11 parameters Parameter Value Slot time σ 9 µs SIFS 10 µs DIFS 28 µs PLCP Header H 24 µs PHY i 6Mbps . . . 54Mbps CW min 15 slots MAC ACK 24 µs Table: Typical IEEE 802.11g parameters Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 8 / 48

Beware of TCP ACKs! The previous model does not take into account the TCP ACKs that go in the uplink direction. TCP ACKs were designed to have low impact on the reverse path (40 bytes against 1500). However, due to the 802.11 overheads, their impact is greater: The effective rate becomes: L C i = T i + TCP ACK i TCP ACK i is the average time to transmit a TCP ACK packet (calculated in a similar fashion), with a typical length of L ack = 40 bytes. Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 9 / 48

The real 802.11g data rates with TCP MAC rates ( C 0 i ) Eff. data rate ( C i ) PHY rates 54 31.9 22.4 48 29.7 21.2 36 24.6 18.5 24 18.4 14.6 18 14.6 12.1 12 10.4 8.9 6 5.57 5.08 Table: MAC rates for the corresponding PHY rates of 802.11g in Mbps. L = 1500 bytes. Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 10 / 48

The real 802.11g data rates with TCP MAC rates ( C 0 i ) Eff. data rate ( C i ) PHY rates 54 31.9 22.4 48 29.7 21.2 36 24.6 18.5 24 18.4 14.6 18 14.6 12.1 12 10.4 8.9 6 5.57 5.08 Table: MAC rates for the corresponding PHY rates of 802.11g in Mbps. L = 1500 bytes. The TCP ACKs may waste 25% of the bandwidth at high rates! Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 10 / 48

The issue of multiple rates From now on we consider only the effective rates C i as given. We will focus on the resource allocation provided by TCP in the presence of these multiple rates . Why are multiple rates an issue? An example... Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 11 / 48

The issue of multiple rates 802.11 AP Point of Presence Class 2 Internet Access Coverage Area Class 1 Class 3 Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 12 / 48

The issue of multiple rates 802.11 AP Point of Presence Class 2 Internet Access Coverage Area Class 1 Class 3 Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 12 / 48

The issue of multiple rates 802.11 AP Point of Presence Class 2 Internet Access Coverage Area Class 1 Class 3 Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 12 / 48

The issue of multiple rates 802.11 AP Point of Presence Class 2 Internet Access Coverage Area Class 1 Class 3 Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 12 / 48

Contents 1 The inefficiencies of TCP over 802.11. 2 TCP resource allocation in a multirate wireless environment 3 A more efficient resource allocation for a single cell 4 A word on more general topologies 5 Implementation and simulations 6 Conclusions Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 13 / 48

Modelling TCP over 802.11 Assume: N stations are downloading data from a single Access Point (AP). Each station i has an effective data rate C i . Data is queued at the AP. The input rate for connection i is x i and the output rate y i . Assuming a fluid model for the queue we have x i y i = � j x j /C j Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 14 / 48

Modelling TCP over 802.11 TCP can be modelled as adapting rate according to some congestion signal (c.f. Srikant, 2004 [9]). x i = k ( x i )( U ′ ˙ i ( x i ) − p i ) where U ( x ) is an increasing and concave utility function, p i is the congestion signal and k ( x i ) > 0 a scale factor. For TCP/Reno like algorithms we can take p i the loss probability. The utility function represents the protocol desire for bandwidth and is typically chosen as U ′ ( x ) = Kx − α with α > 0 a parameter. Example: TCP/Reno takes α = 2 . Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 15 / 48

Modelling TCP over 802.11 To complete the loop, we model the loss probability as the proportion of excess rate. The loss probability for station i becomes: � + � � + � x i − y i 1 p i = = 1 − = p � x i j x j /C j where as usual ( · ) + = max( · , 0) . Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 16 / 48

TCP over 802.11: the downlink model Putting the previous equations together we have: x i ˙ = k ( x i )( U ′ i ( x i ) − p ) , � + � 1 p = 1 − . � j x j /C j which is a Kelly [4, 9] type model of TCP behavior, adapted to the multiple rates scenario of wireless networks. Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 17 / 48

TCP over 802.11: the downlink model Putting the previous equations together we have: x i ˙ = k ( x i )( U ′ i ( x i ) − p ) , � + � 1 p = 1 − . � j x j /C j which is a Kelly [4, 9] type model of TCP behavior, adapted to the multiple rates scenario of wireless networks. Which is the equilibrium of these dynamics? Is this equilibrium globally asymptotically stable? Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 17 / 48

Characterizing the equilibrium Consider the following convex optimization problem: Problem 1 1 � max U i ( x i ) − Φ( x ) C i x i where: �� � x i x i � Φ( x ) = − 1 − log , C i C i i i x i whenever � C i > 1 and 0 otherwise. i Andr´ es Ferragut (Universidad ORT (Uruguay)) NUM in multirate wireless networks RAWNET 2010 18 / 48