Network design considerations 24

ICT4D networks tend to be wireless (Best, 2003; Galperin, 2005). Wireless networks in general, as opposed to wired networks, are low cost and provided instant widespread access. Wireless networks are easily expanded and introduced into situations where no infrastructure already exists (Galperin, 2005). Galperin (2005) mostly discussed 802.11b3, or WiFi (Wireless Fidelity) networks in Latin America. However, WiFi ICT4D networks with broadband uplinks, e.g. with satellite-based Very Small Aperture Terminal (VSAT), are quite common around the world and continue to grow in number (Best, 2003). WiFi is currently the most ubiquitous broadband wireless networking technology because of its standardisation and low cost.

Cellular networks have also provided demonstrable success in digital divide communications (Best, 2003). Cellular GSM (Global System for Mobile communication) and GPRS (General Packet Radio Service) narrowband and 3G broadband connectivity along with up and coming 4G networks like WiMAX (Worldwide Interoperability for Microwave Access) are also becoming more and more prevalent. Many projects leverage standard narrowband GSM networks to provide innovative ICT4D services. Some examples include: SMS reminders for patients to take medication (Anand & Rivett, 2005; bridges.org, 2005c), SMS and GPRS-based agricultural data sharing (Veeraraghavan et al., 2007) and a SIM-based (Subscriber Identity Module) e-commerce system for a rural ICT project (Slay et al., 2007).

A significant difference between WiFi and cellular networks, besides the actual underlying technologies and protocols, lay in who has control of the network and how it is costed. GSM, GPRS and 3G cellular networks are controlled by large telcos, deployed in

3 _{See Best (2003) for a consolidated primer on wireless networking technologies.} 802.11b can achieve 11Mbps throughput at a distance of up to 100m, but is often deployed over much longer distances with consequently less throughput.

controlled spectra and consequently charged based on usage (Best, 2003). WiFi networks, on the other hand, operate in public 2.4GHz and 5.8GHz spectra and tend to be controlled by users (Best, 2003; Purbo 2003, 2004a). To repeat, WiFi networks are standardised, cheap and easy to install (Best, 2003; Galperin, 2005). Once installed, there are no running costs other than maintenance.

WiFi networks also consume very little power and can span large distances (Patra et al, 2007; Sheth et al., 2007; Subramanian et al., 2006; Surana et al., 2007, 2008). There are several WiFi deployment strategies to achieve widespread coverage with a technology designed only for 100m distances: embrace intermittent and fault tolerant networks (Brewer et al., 2005; Du et al., 2006), overcome challenges with low level, e.g. OSI Layer 2, tweaks (CRCNet, 2005; Patra et al., 2007; Sheth et al., 2007), use mesh networks (Bhagwat et al., 2004; Johnson, 2007; Surowiecki, 2006), use WiFi as-is (Galperin, 2005) or some combination of any of these (Best, 2003).

Intermittent connectivity is a common theme in the literature. Pentland et al. (2004) emphasised that users in remote developing regions do not have the luxury of always-on connectivity. Asynchronous services are sufficient to get an ICT4D foothold into a given area (Pentland et al., 2004). Therefore, DakNet is asynchronous and intermittent by design (Chyau & Raymond, 2005; Pentland et al., 2004). DakNet latency is measured in days and hours instead of microseconds and seconds because it uses a data mule4, e.g. bus, motorcycle, bicycle or even an ox-cart fitted with a wireless AP and some data storage. The mobile AP, or MAP, travels along a route and automatically links up with remote wireless access points along the way. In remote areas, APs are situated at a village kiosk (India) or a school (Cambodia). The route of the MAP eventually links up with an existing Internet uplink with which incoming and outgoing transactions to the Internet are enacted. A MAP might drive past a given kiosk or school once a day, and can transfer about 20MB in each direction, with effective goodput5 of 2.47Mbps (Pentland et al., 2004). DakNet relies on short communication distances so no modifications were required to 802.11 protocols or wireless equipment.

4 _{A data mule physically carries a data storage device from one place to another.} 5 _{Goodput is similar to throughput but concerns only application data that gets through.}

TIER designs long-range point-to-point WiFi networks for very challenging environments with standard WiFi equipment. However, these WiFi-based Long Distance (WiLD) networks implement low-level Media Access Control (MAC) layer enhancements (Patra et al., 2007; Sheth et al., 2007) and a Delay Tolerant Networking (DTN) approach that embraces asynchronous data transfer (Demmer et al., 2004; Fall, 2003, 2004; Jain et al., 2004). Asynchronous communication features heavily in the TIER projects, e.g. asynchronous medical tele-consultation (Luk et al., 2008). TIER also aims for more reliable WiLD networks (Subramanian et al., 2006; Surana et al., 2007, 2008). CRCNet also addresses long-range wireless connectivity by tackling low-level OSI Layer 2 protocol timing mechanisms on standard WiFi equipment and protocols (CRCNet, 2005). CRCNet's main objective, however, is to provide synchronous broadband access instead of asynchronous and intermittent access. Regardless of synchrony, the overriding goal for both WiLDNet and CRCNet projects is to design networks that can be managed by local inhabitants. This entails robust software that automatically configures itself and cheap reliable hardware that can be easily installed and maintained by end-users.

Another approach to widespread WiFi access is to use mesh networks (Bhagwat et al., 2004; Johnson, 2007). There are well-founded arguments against the use of long-range mesh networks (Subramanian et al., 2006). However, local short-range mesh networks address many ICT4D challenges. Mesh software can be put into both devices and APs meaning that mesh networks can be expanded easily just by adding more APs and user devices (Bhagwat et al., 2004; Johnson, 2007; Surowiecki, 2006).

While the projects discussed above tackle mostly rural areas in developing regions, MaverickNet is largely an urban phenomenon in Indonesia (Purbo, 2003, 2004a, 2004b, 2005). WiLDNet, DakNet, CRCNet and MaverickNet share a common aspiration in that the end-users should be empowered to set up a network themselves. The MaverickNet strategy is to build easy-to-use, low cost WiFi networks and educate people to deploy and use them (Purbo, 2004a). MaverickNet grew organically to provide Internet connectivity to millions of Indonesians via clandestine and illegal wire line and WiFi links in defiance of Indonesian regulatory legislation (Purbo, 2003, 2004a, 2004b).

In summary of the network design considerations, it should be noted that we infer that the network design for all projects reported herein, e.g. WiLDNet, DakNet, CRCNet and

MaverickNet, is accomplished with the traditional OSI and/or TCP/IP design abstractions. It is telling that stacks are usually not mentioned by name explicitly in publications other than an infrequent reference to an OSI layer, e.g. WiLDNet protocol substitution or CRCNet timing modification at Layer 2. In other words, the use of OSI and TCP/IP stacks is implicit in all projects' network design. Section 3.2.1 will show that the implicit use of traditional stacks also entails an implicit use of QoS to evaluate the efficacy of ICT4D networks.