From: Paul H. Hargrove (PHHargrove_at_lbl_dot_gov)
Date: Wed Dec 01 2004 - 14:38:14 PST
Zoltan,
I think something like the following might work:
len = snprintf(cmd_buffer, cmd_buffer_len, "cr_checkpoint --clobber
--file /proc/self/fd/%d --pid %d", socket_fd, target_pid);
if (len >= cmd_buffer_len) {
/* cmd_buffer too small, deal with it */
exit(1);
}
rc = system(cmd_buffer);
The use of /proc/self/fd/%d will cause the checkpoint to go to an
existing open file descriptor. The --clobber is needed to ensure it
goes straight there rather than going to a temporary file that is then
renamed to the destination (which would fail). Note that this will fail
if the socket descriptor is close-on-exec.
I think it would be easy to implement "--fd <N>" to get the same
behavior as "--clobber --file /proc/self/fd/<N>". In fact, I've added a
request for this feature to our bug database
(http://mantis.lbl.gov/bugzilla) as bug #882.
-Paul
JCDuell_at_lbl_dot_gov wrote:
> Zoltan:
>
> We have not tested checkpointing to a socket, but it should work. We
> are planning to move most of the logic currently in cr_checkpoint into a
> set of library routines, which will allow a socket file descriptor to be
> passed in instead of a regular file. In the meantime, a relatively
> small amount of hacking in cr_checkpoint.c could allow you to do the
> same thing--just add a '--socket <address>' flag, and have it open up a
> TCP socket instead of a file.
>
> Cheers,
>
--
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