From: Paul H. Hargrove (PHHargrove_at_lbl_dot_gov)
Date: Wed Jul 22 2009 - 11:13:05 PDT
Darius,
I am pretty busy, so I'll only give the "outline" of a solution.
What I believe to be the "current best" solution is an "ephemeral
file". In a BLCR callback, one can call cr_get_restart_info() to get a
structure that includes various information. One field of the struct is
the pid of the restart requester (the cr_restart process). Since that
value is unique per-instance it can be used to create a unique file
name. The only trick is that one uses the cr_restart utility instead of
the cr_request_restart() API, one needs to fork(), create the file, then
exec*() rather than just call system("cr_restart ....").
NOTE: in the future we plan to provide an "aux environment" to allow
callbacks to use a getenv-style mechanism to query key-value pairs set
by the restart requester. That will become the recommended mechanism
when available (by SC09 would be great, but I can't promise that).
-Paul
Darius Buntinas wrote:
> I'm trying to figure out the best way to provide a process new
> parameters on a restart.
>
> Take the example of a process which is connected to a server over an
> ephemeral port. The process is checkpointed, and later restarted and
> needs to reconnect to a new instance of the server possibly listening on
> a different port.
>
> When the process was originally started the port information was given
> as a command line parameter or an environment variable. But when it's
> restarted from a checkpoint it gets the original values (as they're part
> of the process image). So I need to find a way to get this new
> information to the restarted process.
>
> One solution would be to use a well-known port rather than an ephemeral
> port. This is not the best solution since there may be more than one
> instance of the server running, and the process needs to connect to the
> correct one. Also, the server may be running on a different node
> altogether.
>
> Another solution is to write this info to a file and have the process
> read the file on restart. Again we'd need a well-known filename which
> would have trouble if there are multiple instances being restarted.
>
> I'm wondering if there is a way to have the restarter process open an fd
> to the server before it forks the new restarted process (I'm not sure
> how exactly this is done) so the restarted process would inherit the
> opened fd.
>
> Any ideas would be greatly appreciated.
>
> Thanks,
> -d
>
--
Paul H. Hargrove PHHargrove_at_lbl_dot_gov
Future Technologies Group Tel: +1-510-495-2352
HPC Research Department Fax: +1-510-486-6900
Lawrence Berkeley National Laboratory