From: Paul H. Hargrove (PHHargrove_at_lbl_dot_gov)
Date: Thu Mar 22 2007 - 11:01:56 PST
Neal Becker wrote:
> I'm doing checkpointing periodically on a simple, single-threaded process,
> like this:
> if (rename (name.c_str(), oldname.c_str()) != 0 and errno != ENOENT)
> die ("rename failed");
> cr_request_file (name.c_str());
> cr_enter_cs(id);
> cr_leave_cs(id);
> if (remove (oldname.c_str()) != 0 and errno != ENOENT)
> die ("remove failed");
>
> I guess this forces the process to wait until the checkpoint is complete. I
> wonder if I can do something more efficient? I'd rather avoid having to mess
> with callbacks, though.
Neal,
What exactly do you mean by "more efficient" here? If you want to
have a backup file, and you want to ensure it is kept until the new file
is complete, I don't see what else you can do but wait for the
completion. The following is an alternative method:
if (rename (name.c_str(), oldname.c_str()) != 0 and errno != ENOENT)
die ("rename failed");
{ char buffer[1024];
snprintf(buffer, sizeof(buffer), "cr_checkpoint -f %s -p %d",
name.c_str(), getpid());
}
if (remove (oldname.c_str()) != 0 and errno != ENOENT)
die ("remove failed");
If you are concerned by the use of "cr_enter_cs(id); cr_leave_cs(id);"
as an idiom for "wait for the checkpoint to finish", then there is not
much I can offer right now. However, we are in the process of designing
and implementing a new checkpoint-request API for inclusion in the 0.6.0
release (summer '07?). That API will have a more explicit "wait for the
checkpoint" call, and that call will block in the kernel, which should
be much "kinder" than the current enter/leave idiom which requires
spinning on an atomic counter in user space.
-Paul
--
Paul H. Hargrove PHHargrove_at_lbl_dot_gov
Future Technologies Group
HPC Research Department Tel: +1-510-495-2352
Lawrence Berkeley National Laboratory Fax: +1-510-486-6900