Re: another laundry or shopping list

[Alan Robertson <alanr@unix.sh>]

carlos wrote:
> 
> Alan Robertson wrote:
> > A process has state internally, and migrating it is very different from
> > restarting it.  You can't migrate a process when a node dies - you can only
> > restart it.  Automatic restart of a process by the OS is probably not a good
> > idea.
> ok, may be some process are not susceptible to restart,but how about
> sending the state in the heartbeats and all the things the process
> need  to start in the last point it left it`s work.

And if it is a process with a 10 megabyte address space, (in general) it's
state is the contents of that 10 megabyte address space plus kernel stack,
etc. ;-)

> For other
> insignificant process like an application designed to run in parallel in
> user space, it may be possible.
> 
> I Know there are a lot of complications that make this problem difficult
> to solve, I agree that starting (or restarting) the process is not a
> process migration. But may  be with a congruent system, with the same
> hour, CPIDS etc ... make it in that way could be possible for a class of
> process.
> 
> doing in that way is probably a bad idea. żin what way can you solve
> that problem in a cluster environment?

One generally has a high-availability management system which tracks what
processes are supposed to be running and what services they're supposed to
be providing and whether they're doing that.

If they're not, the HA system selects where to run the process and starts it
up there.

It is it's job function to know the rules about what can run where and move
them around the cluster according to the rules the administrator (and its
programmers) have given it.

This is normal kind of stuff.  FailSafe does this, heartbeat does this, lots
of other HA software systems do this.

	-- Alan Robertson
	   alanr@unix.sh

Linux-cluster: generic cluster infrastructure for Linux
Archive:       http://mail.nl.linux.org/linux-cluster/