Scyld MPI & debugger

[James Cownie]

> We were wondering what type of debugger works with the Scyld
> implementation of MPI.
> 
> We tried totalview but it crashes and wondered anybody have any
> experience running totalview on scyld.

Totalview will not (yet) work with the Scyld system.

There are a couple of ways in which it could be made to work,
unfortunately, neither is trivial :-

1) The ptrace system call could undergo the same traslations as other
   process control system calls, so that it could be used on the
   front-end "surrogate" processes. In this case all of the processes
   would appear to be local, and no totalview servers would be required.

2) The totalview servers could be run on the slave processors, but
   they'd need to be able to find the "local" pids of the local
   processes, and would need to have ptrace work (and "wait", at
   least), work with those local pids, rather than having them be
   translated back to the global pid space.

The Scyld model is not currently very debugger friendly, since the
multiple pid spaces do not work quite right for the system calls
debuggers need to make :-( (I'm not trying to put down all the good
work the Scyld people have done, and I know that they intend to fix
some of this, I'm just pointing out the reality of the current state
of the world).

Personally I would love to get TotalView working with Scyld Beowulf,
but I have a whole bunch of other TotalView work to do (as do all the
other folks here), so I don't expect to see it happen very soon. (In
case you don't understand "Personally", that means I'm _not_ speaking
for Etnus).

If you want/need TotalView on Scyld, send mail to info at etnus.com and
point out how large a license you'd like to buy. That's what'll make
it happen sooner...

-- Jim 

James Cownie	<jcownie at etnus.com>
Etnus, LLC.     +44 117 9071438
http://www.etnus.com