Newsgroups: comp.parallel.pvm
From: shao@bigtime.cs.colorado.edu (Chung-Shang Shao)
Subject: Re: PVM and distributed loading
Organization: University of Colorado, Boulder
Date: 10 Apr 1995 16:05:22 GMT
Message-ID: <3mbl02$ruf@csnews.cs.colorado.edu>

In article <1995Apr7.072435.24440@aristo.tau.ac.il>,
Yossi Gottlieb <yogo@lune.math.tau.ac.il> wrote:
>: |> It looks like you can spawn processes and assign them to specific machines
>: |> before they start, but I didn't see anything that would allow you
>: |> to move a process from one machine to another after its started running.
>
>: It cannot be done without appropriate operating system support, and
>: few systems provide such support.  In particular, 99% of Unices don't.
>: You are out of luck, I am afraid.
>
>Just a wild idea I've got now -- assuming you have the same binary on both
>systems (identical architecture/os), you dump core (SIGQUIT), take the core
>to the other machine and let it continue.  For purely computational proccesses
>(i.e. those which doesn't talk to various special devices, etc) this might
>even work...  

there are several problems you have to solve:

1. the new machine's pvmd has new host tid, and does not know the
   task which has wrong task id for the new pvmd.
2. the task does not know how to communication with the pvmd,
   since the communication pipes are opened for old pvmd on
   the old machine.

cs shao.


