@IRS 1998–2001

We took part in @IRS: Architecture Intégrée de Réseaux et Services, supported both by the RNRT (Réseau National de Recherches en Télécommunications) and by the CNRS Telecommunications Programme. It was a substantial project, with a total cost of around 20 MF. @IRS set out to develop and experiment with next-generation Internet protocols and the functions that go with them, so as to offer users services matched to their needs whatever the point of access, fixed or mobile, and to take advantage of an increasingly capable and heterogeneous telecommunication infrastructure (ATM, satellite, wireless networks, local networks).

The project focused in particular on the quality-of-service and mobility mechanisms associated with IPv6, the next generation of the IP protocol. The platform built during the project made it possible to evaluate the chosen solutions in the context of new real-time, multimedia and interactive applications. Our contribution addressed resource management in wireless networks to sustain a required level of quality of service, and mechanisms for station mobility.

We defined a QoS architecture coupled with mobility management across a set of 802.11 cells, and designed a micro-mobility scheme at the IPv6 level with fast handoffs between adjacent cells. Quality of service is managed at two levels: within a cell and between cells. The first is handled by the access router, which copes with rapid changes in local conditions: mobile hosts tell the access router how much bandwidth they need, and the router configures their QoS parameters. The second concerns a set of cells attached to a border router, which sets long-term policies for the access routers.

The group #

Scientific lead:

• Andrzej Duda, professor at INPG

Permanent staff:

• Gilles Berger-Sabbatel, CNRS research scientist
• Franck Rousseau, associate professor at INPG

PhD students:

• José Antonio García Macías
• Leyla Toumi

The problem #

LSR’s part in @IRS concerned mainly quality of service and mobility in wireless networks under IPv6: assessing how well wireless networks can support quality of service, and proposing, testing and evaluating mechanisms that improve it for nomadic stations attached to the wireless network.

Wireless networks rest on a set of devices, called base stations, relaying transmissions between a fixed communication infrastructure and the mobiles. The area a base station covers forms a cell. Wireless networks therefore raise a number of specific problems:

• It is hard to guarantee the availability of resources on a radio channel whose access method does not necessarily allow fine enough control of the data flows. In particular, flows originating from mobile stations can only be controlled on the stations themselves, in a distributed way. Flows towards the mobiles, by contrast, admit more precise control.
• The bandwidth actually available to a mobile can fluctuate as the station moves, and also with the traffic the cell is carrying for other mobiles.
• Movement may take a mobile station out of a cell’s coverage. The station then has to register with another cell and routing has to change accordingly. This may happen mid-communication, and should disturb transmissions and quality of service as little as possible — in reliability (no data lost because of the cell change), in delay (the change must not introduce more delay than the application can accept), and in bandwidth (the bandwidth the application needs must remain available in the new cell).

We chose to restrict the study to wireless local networks. The only longer-range networks usable at the time were wireless telephony networks, whose low bit rate (9.6 kbit/s) seriously limits their use for data, and where the wireless network presents itself as a black box inside which no control is possible.

Transmission technology #

The technologies available for wireless local networks were essentially IEEE 802.11 and Wireless ATM. IEEE 802.11 was then the most mature and widespread, and among its implementations we selected Lucent’s WaveLAN.

WaveLAN transmits in the 2.4 GHz band using DSSS (Direct Sequence Spread Spectrum). The channel is shared using CSMA/CA (Carrier Sense Multiple Access with Collision Avoidance). It offers 2 Mbit/s, dropping to 1 Mbit/s when transmission quality degrades (a falling signal-to-noise ratio). WaveLAN Turbo cards can in fact reach 6 Mbit/s, though that falls outside the IEEE 802.11 standard. WaveLAN’s range can reach 300 m outdoors, but is limited to under twenty metres indoors with partitions.

WaveLAN supports two modes of use: ad hoc networks, where mobiles communicate directly with each other and need not be attached to any fixed infrastructure; and infrastructure networks. In the latter, mobiles communicate through base stations called WavePoints, themselves attached to a conventional Ethernet network for which they act as bridges. Cells attached to the same Ethernet are seen as belonging to the same network at the IP level.

The interfaces of mobile stations continuously assess the signal-to-noise ratio of the transmission received from the base station they are attached to. When it falls below a threshold, the interface enters a mode in which it looks for other base stations offering better transmission quality. When a further threshold is reached, the cell change is triggered: the mobile registers with a new base station and transmits through it. This mechanism is called handoff. In the general case mobiles stay within the same IP network, so there is no mobility at the IP level.

Quality of service #

The IETF proposed two approaches to quality of service: integrated services (IntServ) and differentiated services (DiffServ). Both have limits. IntServ rests on a signalling protocol, RSVP, considered not to scale and so ill-suited to large networks. DiffServ tries to sidestep the difficulty by reducing the problem to managing traffic classes with different priorities, combined with policing of sources. It is, however, hard to characterize the quality of service obtained and to guarantee that, for a given network configuration, it matches what applications ask for. Because of these difficulties the field remains open, work on both integrated and differentiated services continues, and new approaches keep being proposed. We therefore set out to evaluate some of them.

Wireless local networks bring several difficulties of their own: a MAC layer with no notion of priority as it stands, hosts that move, and channel bandwidth that fluctuates with the ambient electromagnetic noise. In practice, though, a properly dimensioned network can be made to guarantee a minimum bandwidth.

Within @IRS we set out to use the mechanisms proposed at the IP level to implement quality of service. That requires controlling the admission of mobiles to each cell. Cells are therefore managed through a router controlling access and making it possible to implement priorities in traffic forwarding, and the IP stacks of mobile hosts have to include source policing (token bucket). This adds a difficulty to handoff, which may then involve a change of network and so require routing changes. We set out to find mechanisms suited to mobility local to a single administrative entity (subnets within one local network), allowing a more efficient handoff than the mechanisms proposed by Mobile IP — in particular, limiting the delay handoff introduces and the risk of data loss.

Choice of architecture #

In the architecture we selected, the wireless network attaches to the rest of the experimental network through a border router, which performs:

• basic functions: classification and marking, policing, admission control;
• specific functions: resource management, handover.

Three approaches to adding quality-of-service functions are possible:

• At the MAC level, by modifying the access point: this would mean reimplementing handover using a resource management function provided by the border router. Short of close collaboration with the hardware vendor, the scope for such a change looks rather limited. IEEE 802.11 offers no support for quality of service, though Lucent has made proposals in that direction.
• At the MAC level, by designing an ad hoc access point: here one exploits the direct station-to-station communication the network allows, rewriting the bridging (Ethernet to WaveLAN) and cell management functions and folding quality-of-service functions into them.
• At the IP level: the WaveLAN access point attaches to a PC router over a dedicated Ethernet segment. This solution costs more in hardware complexity, requires managing an IP subnet per WaveLAN cell and handling handover at the IP level. But it also lifts quality-of-service handling to the IP level, where the available mechanisms can be used independently of the MAC layer, and it avoids depending too closely on what the hardware offers and on how well documented that is.

Resource management can be handled as follows:

• bandwidth-broker functions performed by the border router, reserving part of the resources (throughput) for a class of service;
• the cell manager (the access-point router) limiting the number of mobiles of a given class per cell, and allowing (overlapping) cells to be reserved for certain classes of service. It performs a token-bucket function under the control of the border router.

Experimental platform #

The planned experimental platform consists of:

• a PC router providing the ATM interconnection with Renater-2 and the border router functions, with a VP to the core router at INRIA Sophia Antipolis;
• two WaveLAN base stations communicating with the mobiles, attached to the PC router over Ethernet;
• several mobile machines (at least two laptops).

Several options are possible for the base stations:

• Use the manufacturer’s base stations (WavePoint) connected directly to the Ethernet. This allows no control over resource usage, and handover is handled at the MAC level.
• Use fixed PCs fitted with WaveLAN cards. This requires reimplementing all the WavePoint functions. It offers the widest range of possibilities, at the cost of considerable implementation work, and with no guarantee in advance that the card vendor will supply all the necessary information.
• Use WavePoints connected to a PC router over a dedicated Ethernet segment. This costs more in hardware, but gives better control over data flows, particularly at the IP level, without having to touch the lower communication layers.