Re: Clusterwide pids

[Bruce Walker <bruce@kahuna.cag.cpqcorp.net>]

> 
> On 20010711 irbis@orcero.org wrote:
> 
> I have read sometimes about 32bit pids in lkml. I think the first thing to do is
> ask lkml (well, really Linus) what are the plans for 32bit pids on main kernel
> stream. Perhaps it is a 2.5 planned feature.

In 2.4, all the pids in the kernel are 32 bit and getpid() returns a
32 bit number.  We have been node-encoding (i.e. CPIDing) for 9 months
and have noticed only 1 minor problem in Red Hat 7.1.  What libc 
converts a 32 bit pid from the kernel to a 16 bit pid?  B.t.w., top and
ps seem to work just fine with CPIDs.

If in fact the size of the pid is not an issue and the code to create
CPIDs is clearly not an issue, I think the key discussion point would
be  how and where to designate a node number.  For the Cluster Infrastructure
project (CI), we originally were setting the node number in lilo but
on Linus's suggestion, we have made cluster enabling and joining 
command driven.  That means you don't have a node number until you are
ready to join a cluster, potentially long after some processes have
been created.  For the open SSI cluster, the initialization is done
from commands in the ram disk so a node number is known very early.

Given a scheme to determine the node number early in boot, I would suggest
that 0 be the default (no clustering; not a valid node number) and
you assign a node number if you intend to cluster at some point.  Clearly
we don't want to hard code the node number into the kernel.
> 
> > Not necessarly. In fact, the only working SSI fully transparent migration
> >scheme running today on productions environments -Mosix- has no
> >a common PID space.
> >
The Linux version of Mosix has SSI transparent migration but is not
an SSI cluster.  Clusterwide pids are more important in an SSI
cluster.  The distinction is that in Linux version of Mosix, processes
that migrate away from their home node still see the view of their home
node. Each home node has a different view of processes, devices,
IPC objects (fifos, unix sockets, messages queues, etc), filesystems,
networking, etc.).  In an SSI cluster, all processes on all nodes see the same view
of all resources, including all other processes.

bruce walker
Open SSI Cluster architect
Compaq Computers

> -- 
> J.A. Magallon                           #  Let the source be with you...        
> mailto:jamagallon@able.es
> Mandrake Linux release 8.1 (Cooker) for i586
> Linux werewolf 2.4.6-ac2 #1 SMP Sun Jul 8 23:57:11 CEST 2001 i686
> 
> Linux-cluster: generic cluster infrastructure for Linux
> Archive:       http://mail.nl.linux.org/linux-cluster/


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