Newsgroups: comp.parallel.pvm
From: wjiang@thud.cs.utk.edu (weicheng jiang   )
Subject: Re: Patch #8 for PVM version 3.3 is available on Netlib
Organization: Computer Science Dept, University of Tennessee, Knoxville
Date: 7 Aug 1995 21:00:35 GMT
Message-ID: <405utjINNhkj@CS.UTK.EDU>

In article <3vq3gj$l2k@sunsystem5.informatik.tu-muenchen.de>, stellner@Informatik.TU-Muenchen.DE (Georg Stellner) writes:
|>
|> Another interesting question would be, how the implementation of PVM on
|> top of MPI handles dynamic process creation with pvm_spawn. The MPI 
|> standard definition doesn't include process creation. Or does it mean,
|> PVM ontop of MPI doesn't offer dynamic processes?
|>

This version of PVM is fully compatible with the generic version of PVM in
the sense that the other MPP ports are. When a PVM program spawns a group
of tasks, those tasks will belong to the same MPI_COMM_WORLD. When a second
group of tasks are spawned, they'll belong to a different MPI_COMM_WORLD.
Tasks in different MPI_COMM_WORLD can communicate via pvmd over TCP sockets.
So message-passing within the same MPI_COMM_WORLD is a lot more efficient.
A user should minimize communications between tasks not spawned by the same
pvm_spawn().

-Weicheng Jiang
PVM Group, University of Tennessee


