Re: Clusterwide pids

["David L. Nicol" <dnicol@cstp.umkc.edu>]

Patrick Spinler wrote:
> 
> "Albert D. Cahalan" wrote:
> >
> > It is better to pass out chunks of PID space or
> > lists of free PIDs. 
...

> This scheme is quite straightforward, which is good. But in my limited
> understanding it seems to place a cluster architechure in one of two
> possibly difficult to resolve scenarios:
> 
> a) cluster_id's and pid ranges are hardcoded at boot time
>    -) potentially difficult to administer large or highly dynamic
> cluster(s)
> 
> b) node has to join cluster and request pid range's & cluster id before
> starting any processes
>    -) all key cluster membership management & control must be kernel
> level
>       (this may be necessary anyway)
>    -) still have to assume certain processes & pid's are local node
> only,
>       e.g. init is assumed to be pid 0 in many places, correct ?
> 
> How might you address these problems ?  Have I misunderstood ?
> 
> -- Pat

Reserve a range (<8K, for instance) for nonmigratable on any machine,
then divide up the higher PID numbers.  complex machines still might
need translation tables, but trivial clusters wouldn't any more.  
B*-like division of PID space could then be static or dynamic.

Early Beowulf (bproc?) used hard PID division, then later switched
to translation tables, as I vaguely understand it.  Anyone reading
this care to explain why, or post an URL to the discussion?

-- 
                                           David Nicol 816.235.1187
                      Irish Government Warning: SMOKERS DIE YOUNGER


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