Related Work

In this section we review the main mobility proposals.

Several link layer technologies provide mobility at the link layer (e.g., as in Ricochet [7], 802.11b, or GSM). However, these solutions preclude mobility across link layer technologies. In addition, hiding mobility at the link layer results in a reinvention of mobility support in each new wireless system; solving the mobility problem at the network layer results in a reusable mobility infrastructure for all link technologies.

One proposal to achieve mobility in the Internet is Mobile IP (MIP). MIP in IPv4 [1] and IPv6 [8] uses an explicit indirection point, called the Home Agent (HA), to encapsulate and relay the Correspondent Host's (CH) initial packet to the Mobile Host (MH). MIP provides the following options that determine how the following packets are routed: 1) triangle routing, 2) bidirectional tunneling, and 3) route optimization.

As noted by Cheshire and Baker [9] no MIP routing option is clearly better than the others; instead, different options are suitable for different circumstances. Options (1) and (2) preserve location privacy, but routing can be inefficient when the MH and CH are close relative to their distance from the HA. With route optimization (an extension in MIPv4 [10], but standard in MIPv6), the MH conveys its care-of IP address to the CH using a Binding Update (BU). Routing is efficient because the ratio of the latency of the optimized route to the latency of the shortest IP path (or latency stretch) is 1.0. However, the CH must be modified to support MIPv4 with route optimization or IPv6. This also exposes the MH's current care-of address (and therefore its location) to the CH, thus compromising location privacy. In certain delay-sensitive or real-time applications, the latency involved in handoffs can be above the threshold if the MH is far away from the CH.

In general, the dependence in MIP on a fixed HA reduces fault tolerance. If the HA or its network fails or is overloaded, then the MH will be unreachable.

To address routing anomalies and robustness issues associated with a fixed HA, researchers have proposed the notion of dynamic home agents in MIPv4 [11]. However, the actual algorithm used to discover and allocate a nearby home agent is still under investigation. MIPv6 provides a dynamic home agent address discovery mechanism [8] that allows a MH to dynamically discover the IP address of a HA on its home network. This scheme increases the robustness of MIPv6 as the HA is no longer a statically fixed entity, but it does not address routing inefficiencies caused by routing through the HA when the MH is far away from its home network.

Recently, two mechanisms have been proposed to increase handoff performance in MIPv4 and MIPv6: (1) low latency handoff [12], and (2) fast handover [13]. The first mechanism attempts to send a BU in advance of an actual link-layer handoff when the handoff is anticipated. However, timing must be arranged such that the BU completes before the actual handoff does, which may be hard to achieve in practice. Similar in concept to Regionalized Tunnel Management [14] and Hierarchical Mobility [15] extensions in MIPv4 and MIPv6, the second mechanism sets up a bi-directional tunnel between an anchor Foreign Agent (FA) that stays the same during rapid movements and the current FA. This allows the MH to delay a formal BU to the HA which minimizes the impact on real-time applications. However, this mechanism relies on the existence of a FA in each network the MH visits. Furthermore, the use of link-layer triggers and inter-FA advertisements in these mechanisms assumes a homogeneous link-layer technology. In contrast to both these mechanisms, ROAM supports fast handoff by giving end-hosts implicit control over trigger placement.

The Host Identity Protocol [16] supports mobility by decoupling the transport from the network layer, and binding the transport to a host identity. Similarly, Location Independent Networking for IPv6 [17] specifies a unique identifier for a host regardless of its location. These ideas are in line with the seminal work by Jerome Saltzer on host identifier and locator separation [18]. However, $i3$provides a general-purpose indirection infrastructure that enables a variety of communication services beyond mobility, such as multicast, and transcoding. Section 4.2.2 will present the use of $i3$multicast to support soft handoffs for mobile hosts.

Supporting Mobility for TCP with SIP [19] spoofs constant TCP endpoints in a similar way to MIP with route optimization. This requires modifying the IP stack of the CH.

The Mobility Support using Multicasting in IP (MSM-IP) architecture [20,21] implements mobility using IP Multicast [22]. The main advantages of MSM-IP are that it can have low latency and do handoffs with little or no packet loss. Several studies [21] [23] [24] have shown that multicast mobility can cut the latency stretch of Mobile IP in half and significantly reduce packet loss due to handoffs. However, the MSM-IP location service is a single point of failure and is vulnerable to overload, network faults, and host faults.

In TCP Migrate [2], both the MH and CH use a modified form of TCP which can tolerate a change in IP address during a connection. The CH uses DNS to learn the current address of the MH, which updates DNS every time it moves. Since TCP Migrate does not use an indirection point, it can achieve an optimal latency stretch of 1.0 and is as fault tolerant as IP routing. However, it lacks simultaneous mobility support, requires modification of the TCP implementations on both the MH and the CH, and does not preserve location privacy. TCP Migrate is well suited for person-to-server applications with short-lived flows like email and web browsing.

The mobility schemes previously described in this section track mobile hosts. In contrast, personal and session mobility schemes (e.g., The Mobile People Architecture (MPA) [4] ICEBERG [5], and Telephony Over Packet networkS [3]) track people or sessions. This allows redirection of new sessions or migration of active sessions to a completely different application or device according to user connectivity (e.g., which devices are currently accessible to the user) and user preferences (e.g., less expensive or higher performance). In contrast, Mobile IP redirects flows to the same device regardless of whether the user can actually use the device or not. The costs of personal/session mobility schemes are modifications to applications (unlike Mobile IP) and an indirection infrastructure (e.g., the Personal Proxy in MPA).

In contrast to all of the above schemes, the novelty of our approach is the use of an overlay indirection infrastructure that gives endhosts control over the placement of the indirection points. As a result, ROAM achieves efficiency, robustness, location privacy, and simultaneous mobility. In addition, the flexibility of $i3$identifiers allows ROAM to support mobility at any layer; $i3$ identifiers can be bound to hosts as well as sessions and people.

Figure: (a) $i3$'s API. Example illustrating communication between two nodes: (b) Receiver $R$ inserts trigger $(id, R)$. (c) Sender $S$ sends packet $(id, data)$.