Connecting to the Pvmd<A NAME=1247> </A>



next up previous contents index
Next: Protocols Up: Libpvm Library Previous: Language Support

Connecting to the Pvmd 

 

On the first call to a libpvm function, pvm_beatask() is called to initialize the library state and connect the task to its pvmd. Connecting (for anonymous tasks) is slightly different from reconnecting (for spawned tasks).

The pvmd publishes the address of the socket on which it listens in /tmp/pvmd.uid, where uid is the numeric user id under which the pvmd runs. This file contains a line of the form 7f000001:06f7 or /tmp/aaa014138

This is the IP address and port number (in hexadecimal) of the socket, or the path if a Unix-domain socket. To avoid the need to read the address file, the same information is passed to spawned tasks in environment variable PVMSOCK.

To reconnect, a spawned task also needs its expected process id. When a task is spawned by the pvmd, a task descriptor is created for it during the exec phase. The descriptor must exist so it can stash any messages that arrive for the task before it reconnects and can receive them. During reconnection, the task identifies itself to the pvmd by its PID. If the task is always the child of the pvmd (i.e., the exact process exec'd by it), then it could use the value returned by getpid(). To allow for intervening processes, such as debuggers, the pvmd passes the expected PID in environment variable PVMEPID, and the task uses that value in preference to its real PID. The task also passes its real PID so it can be controlled normally by the pvmd.

pvm_beatask() creates a TCP socket and does a proper connection dance with the pvmd. Each must prove its identity to the other, to prevent a different user from spoofing the system. It does this by creating a file in /tmp writable only by the owner, and challenging the other to write in the file. If successful, the identity of the other is proven. Note that this authentication is only as strong as the filesystem and the authority of root on each machine.

A protocol serial number (TDPROTOCOL) is compared whenever a task connects to a pvmd or another task. This number is incremented whenever a change in the protocol makes it incompatible with the previous version.

Disconnecting is much simpler. It can be done forcibly by a close from either end, for example, by exiting the task process. The function pvm_exit() performs a clean shutdown, such that the process can be connected again later (it would get a different TID).



next up previous contents index
Next: Protocols Up: Libpvm Library Previous: Language Support