Startup Phase

This section reports our experience with implementing the MPI startup phase (MPI_Init()) under U-Net on Fast Ethernet. We encountered and overcame several difficulties which we feel are worthwhile mentioning.

Setting up a U-Net channel to a remote host requires knowledge of

1. the local U-Net port
2. the remote host's MAC address
3. the remote host's U-Net port

• The master queries its MAC address from the U-Net Tulip device, and reserves a set of ports, one for each channel to a slave.
• The master reads the names of the slaves from the host file, and starts up the slaves with rsh, passing its host name as an argument.
• When the slaves come up, they also learn about their MAC address and reserve a set of ports to communicate to all peers.
• The slaves set up a TCP connection to the master and tell it their MAC address and the reserved port numbers.
• Now the master has complete information about MAC addresses and port numbers at each slave, and it sends out the relevant information to the slaves.
• At that point, the master and all slaves know the MAC address and corresponding port numbers for their peers, and they set up the U-Net channels.
• Master and slaves synchronize via another TCP two-way handshake to make sure that all U-Net channels are set up completely before returning from MPID_Init(). This ensures that no communication is attempted before all U-Net channels are set up completely.

One could also design a faster startup procedure, where the slaves set up U-Net channels with the master in a star topology early on, and then use the U-Net channels to communicate the port numbers. We decided to use TCP because first the startup phase is not performance critical, and second we avoid dealing with reliability issues.