The libpvm function pvm_mcast() sends a message to multiple destinations simultaneously. The current implementation only routes multicast messages through the pvmds. It uses a 1:N fanout to ensure that failure of a host doesn't cause the loss of any messages (other than ones to that host). The packet routing layer of the pvmd cooperates with the libpvm to multicast a message.
To form a multicast address TID (GID) , the G bit is set (refer to Figure ). The L field is assigned by a counter that is incremented for each multicast, so a new multicast address is used for each message, then recycled.
To initiate a multicast, the task sends a TM_MCA message to its pvmd, containing a list of recipient TIDs. The pvmd creates a multicast descriptor (struct mca) and GID. It sorts the addresses, removes bogus ones, and duplicates and caches them in the mca. To each destination pvmd (ones with destination tasks), it sends a DM_MCA message with the GID and destinations on that host. The GID is sent back to the task in the TM_MCA reply message.
The task sends the multicast message to the pvmd, addressed to the GID. As each packet arrives, the routing layer copies it to each local task and foreign pvmd. When a multicast packet arrives at a destination pvmd, it is copied to each destination task. Packet order is preserved, so the multicast address and data packets arrive in order at each destination. As it forwards multicast packets, each pvmd eavesdrops on the header flags. When it sees a packet with EOM flag set, it flushes the mca.