Multicast-based Soft Handoff

As a result of multicast-based soft handoff, the $i3$server will send duplicate encapsulated packets to the MH. To prevent the MH's TCP/IP stack and applications from receiving duplicates of the inner packet (i.e., original IP packet, see Figure 7), the ROAM proxy suppresses duplicates during multicast-based soft handoff.

The ROAM proxy maintains a small window of MD5 [30] digests of recent packets. The proxy computes digests over the first 20 bytes of the IP header and the first 8 bytes of the transport header. The first 28 bytes of a packet are sufficient to differentiate non-identical packets in practice [31]. To minimize duplicates, the window size must be sufficiently large so that a duplicated packet that arrives both via a very low latency link and via a very high latency link will be caught in the window. We use a window size of 1 second, which should be sufficient even if one path is very congested or contains a 500ms satellite link. Note that this scheme detects duplicates received on different addresses, which means that only duplicates generated by $i3$are eliminated, and not duplicate packets sent by the sender (e.g., TCP dup-ack). We show in Section 7 that a TCP bulk transfer flow using multicast-based soft-handoff achieves similar throughput to a flow without mobility.

Another implementation issue is when does the proxy stop using multicast. Our algorithm removes an address when a large fraction of its packets are duplicates as this indicates that the address is redundant. The ROAM proxy maintains a counter of duplicate packets received on both addresses ($d$) and a counter of packets received on each address ($p_i$). When $d > \min(p_0/k, p_1/k)$, we simply remove the address that has received fewer packets in the last window. The value $1/k$ is a constant indicating the fraction of an address's packets that must be duplicates before the address can be dropped. In addition, the proxy uses a timeout $t$ to prevent a newly added address with poor connectivity from being removed until the timeout expires. In our implementation, we use $k=2$ and $t=30$ s.