Documente Academic
Documente Profesional
Documente Cultură
11
Wireless Networks
Henrique Moura , Gabriel V. C. Bessa , Marcos A. M. Vieira , Daniel F. Macedo
Computer Science Department
Universidade Federal de Minas Gerais, Brazil
Emails: {henriquemoura, gabrielvcbessa, mmvieira, damacedo}@dcc.ufmg.br
AbstractWireless Networks have become ubiquitous and advocates the possibility of: (i) more evolvable networks, as
dense to support the growing demand from mobile users. To we should be able to implement new features without changing
improve the performance of these networks, new approaches are the platform; and (ii) greater network efficiency, since more
required, such as context and service aware control algorithms,
which are not possible on todays closed proprietary WLAN advanced (context and application aware) algorithms can be
controllers. In this work, we propose Ethanol, a software- easily deployed (usually in run-time).
defined networking architecture for 802.11 dense WLANs. This However, in existing SDNs (e.g. OpenFlow), the data plane
paper describes the benefits of programmable APs, and proposes is programmable only on switching elements. We advocate that
Ethanol, an architecture for network-wide control of QoS, user
SDN-enabled access points will be able to improve the man-
mobility, AP virtualization, and security on 802.11 APs. The
proposal is evaluated on a prototype using off-the-shelf APs over agement of WLANs even further than proprietary controllers.
three use cases. Having an open API will allow the deployment of context
and application aware control algorithms, an impossible task
I. I NTRODUCTION on current wireless controllers. Further, an open API for access
Wireless Networks have become ubiquitous and dense. In points will allow rapid innovation, accelerate the deployment
a recent study [1], more than 90 percent of U.S. colleges of new network services and foster the competition among
intend to increase their wireless infra-structure to support the vendors. In the same way that SDN has generated many
demand of mobile applications. The same trend is expected innovative applications on the wired domain, we believe that
in home networks. With the Internet of Things, more devices SDN-enabled access points will accelerate the innovation on
will connect to the network. Cisco predicts that 50 billion WLAN management and operation.
things will connect to the Internet by 2020 [2]. At the same This paper proposes an SDN architecture for dense IEEE
time, emerging high-speed network standards (e.g. 802.11ad) 802.11 Wireless LANs called Ethanol. Ethanol refactors the
are migrating to higher frequencies, which are absorbed by control plane functionalities between the APs and the con-
walls, thus requiring the deployment of multiple access points. troller. The centralized controller actuates on APs, controlling
Thus, future networks must be optimized to increase their features such as quality of service, client mobility and asso-
scalability and effectiveness in dense environments through ciation/disassociation, link-level parameters and current state,
a more refined control of all the devices in the network. and virtual APs. The proposed architecture was prototyped on
Current management architectures for WLANs employ pro- commodity home routers, and the performance of Ethanol was
prietary controllers, which perform network-wide optimiza- evaluated on three case studies involving quality of service, the
tions such as the selection of best operational channels, the control of client association, and ARP packet suppression.
adjustment of transmission power at each AP, faster client The remaining of the text is organized as follows. Sections II
mobility and enhanced traceability, as well as security and QoS and III provide a background on SDN and the challenges of
policy enforcement. However, those controllers only manage SDN on wireless networks, respectively. The Ethanol architec-
compatible devices, usually of a single manufacturer, since ture and its components are described in Section IV, followed
they rely on proprietary interfaces and proprietary MIBs. Fur- by its implementation in Section V. The implementation and
ther, current wireless network controllers for WLAN remain the results obtained on a testbed are presented in Section VI.
closed and (mostly) proprietary [3], and the industry has little The related work is discussed in Section VII. Finally, Sec-
incentive to change. tion VIII presents our conclusions and future work.
Software Defined Networking (SDN) is a recent network
paradigm that restructures computer networks by separating II. BACKGROUND ON SDN
the control plane from the data plane, bringing dynamic
programmability to the networks [4]. SDN permits network In this section, we provide an overview of the existing
elements to run simple programs, and is said to reduce the control methods for SDNs and for wireless networks.
complexity of network configuration and network management SDNs are characterized by the existence of a control system
while providing leverage to innovation in a simpler way. SDN (software) that controls the switching fabric of routers and
F. User Location
Fig. 1. Ethanol Access Point Implementation
The location of the user is an important element for location-
aware services. [11] show that indoor localization is possible Figure 1 depicts the Ethanol AP architecture. An Ethanol
with no pre-deployment effort using the EZ localization algo- AP has two parts: wired and wireless components. A wired
rithm. This algorithm requires as input that stations perform component is a configurable switching element that supports
RSS (Received Signal Strength) measurements of known APs the OpenFlow protocol. Since OpenFlow does not provide a
or a location fix (e.g. a GPS lock near doors or windows). control interface for QoS and for wireless components, the
RSS readings can be requested to the stations using 802.11k. Ethanol Agent adds this functionality to the AP, using an
Fig. 2. Ethanol control API model. Green entity indicates current OpenFlow model. Elements in red indicate extensions to OpenFlow.
commodity wireless access points, connected to the controller Suppose that a dense network has two access points, both
using ethernet. The chosen model was a Broadcom WRT54GL covering the same location, in order to accommodate a larger
router running OpenWRT, a linux distribution for embedded number of clients. Those APs are allocated to non-interfering
devices. Its hardware provides one 802.11b/g Wi-Fi interface channels, and the objective is to assign the same amount of
and a 4+1 100Mbps Ethernet switch. We installed Pantou, a clients to both APs for load-balancing. Nowadays, clients asso-
userspace OpenFlow router that supports version 1.0 of the ciate based on signal strength. So, if an AP has a slightly better
specification. signal, most clients will attempt to connect to the AP with
Ethanol exploits the QoS elements of Pantou and Linux. the strongest signal, generating a load imbalance. Commercial
Pantou uses Hierarchical Token Bucket (HTB) implementation wireless controllers use proprietary algorithms that rely on
as its main QoS mechanism. HTB is a packet scheduler that generic network behavior to control the association process.
supports priority queuing as well as maximum and minimum The network administrator can only set a few parameters, but
rate allocations. Ethanol can control outbound traffic on a link cannot define the behavior he/she really wants.
using HTB by associating each flow to a queue. We also In Ethanol, the controller is notified whenever new asso-
modified the Linux wireless subsystem in order to control ciation attempts occur, as shown in Algorithm 1. So, the
client association. In Ethanol, the AP first sends a message controller is able to deny the association or redirect it to other
to the controller asking whether a client should be allowed to APs in order to perform load balancing (using 802.11r), or
associate. using application-specific rules. In this experiment we imple-
VI. E XPERIMENTS ment a load-balancing scheme where the controller divides
This section describes the experiments performed on the the number of clients equally among APs using only the
prototype. disassociation feature. Each association request is sent to the
controller, which decides if the client should be allowed to
A. Load-Aware Client Association associate to the AP based on the actual number of stations
One way to improve the throughput of dense wireless already associated (the fewer the better). Observe that Ethanol
networks is to control which client is associated to each AP. does not need to modify the clients software to control
5: macsrc packet.macsrc
5
6: IPsrc packet.IPsrc
4
7: macT oP ort[macsrc ] event.port . maintains macToPort mapping
HTB min rate = 3M 8: arpT able[macsrc ] packet.IPsrc . updates ARP table
3 9: if packet.type! = ARP then . all non ARP packets must be transmitted
10: (new F low(event)).apply()
2 11: else . treat every ARP packet
HTB min rate = 1M 12: if !f indIP (arpT able, IPdst ) then
C2 starts transmitting
1 13: f lood(packet) . dont know destination, then flood
14: else . if destination is known, then:
0 15: if packet.arp = REQU EST then . (a) respond to the sender
50 100 150 200 16: SendArpReply(macdst , macsrc , macT oP ort[macsrc ])
Time (sec)
17: else . (b) transmit response directly to the destination
18: SendArpReply(macsrc , macdst , macT oP ort[macdst ])
19: end if
Fig. 4. Controlling traffic to a server 20: end if
21: end if
22: end function
C. ARP Overhead
Cheng [5] et. at. analyzed traces of a WiFi campus network. wireless client. We configured the controller to switch the ARP
They concluded that ARP (Address Resolution Protocol) pack- filter algorithm on and off.
ets consume almost 10% of the air time of wireless links. Figure 5 shows the total amount of traffic (inbound and
ARP traffic is sent in broadcast, and all ARP broadcasts from outbound) as well as ARP traffic on the wireless link. When
the wired network are also sent in broadcast on the wireless the controller is not filtering ARP packets, the wireless client
channel. Moreover, this traffic scales with the user population, receives those packets on its wireless interface. Most of those
while the wireless capacity remains constant. packets are wasting bandwidth, except for the few requests
Since the controller has knowledge of the IP and MAC directed to the wireless client. On the first 120 seconds, we see
mapping, the controller can solve the ARP requests. Thus, ARP traffic on the wireless side. When the controller starts to
by filtering ARP traffic on the APs, it is possible to free up manage the ARP Requests (gray area on the figure), the ARP
to 10% of the air time. traffic goes almost to zero, because the controller knows the
In this experiment we avoid unnecessary ARP broadcasts wireless clients MAC address. The actual traffic to and from
by controlling the transmission of ARP messages using Open- the wireless client keeps the table up to date. The only ARP
Flow. The benefit of Ethanol here is the fact that wired and traffic seen on the wireless medium when the ARP limiting
wireless ports of the AP are controlled using OpenFlow. algorithm is active is requests from the wireless client, and the
Commercial wireless controllers provide solutions based on first request to this client.
proxy ARP that partially address the problem shown in [5].
Ciscos controller, for instance, can act as a proxy ARP to the 90
ARP trac
network requests, generating ARP replies, however the replies 80 Other trac
are broadcast to the entire network, and this approach does not 70
50
the controller, in such case ARP packets are not handled.
40
On this experiment the Ethanol controller captures all ARP
30
packets and performs simple matches on its internal ARP table.
This behaviour is captured by Algorithm 3. If the controller 20
does not recognize the IP address, it floods the request to all 10 ARP control o ARP control on ARP control o