Conference turris::digital_unix

I heard that there is, in fact, such a feature in the free BSDs.
Something along the lines core.`base-name-of-the-executable`.$$
(in shell notation here but handled by the kernel). If you feel
strongly enough for this feature I guess you could QAR it?

-- 
Jarkko Hietaniemi Digital/Finland MCS/OSS DTN 879-4738
[email protected]   +358-(0)9-434 4738

[Posted by WWW Notes gateway]

> I heard that there is, in fact, such a feature in the free BSDs.
> Something along the lines core.`base-name-of-the-executable`.$$
> (in shell notation here but handled by the kernel). If you feel
> strongly enough for this feature I guess you could QAR it?

	Of course the obvious problem with turning on such a
	feature system-wide is that all your disc space can
	fill up with core files if you're not careful.

    Yep, we'd probably want to rig it up to work something like
    the uac command works for unaligned access fixup/reporting.

    RE: .3
    
    >        Of course the obvious problem with turning on such a
    >        feature system-wide is that all your disc space can
    >        fill up with core files if you're not careful.
    
    You mean like version numbers?  :-) :-)
    
    				-John

    set file core /version_limit=5

    % set file core /version_limit=5
    set: Variable name must begin with a letter.
    
    Rats!

    I forgot to mention steps 1 and 2 (1: add version numbers to file
    system, and 2: write DCL shell).

> I forgot to mention steps 1 and 2 (1: add version numbers to file
> system, and 2: write DCL shell).

	Well I could see adding version numbers, but screw DCL (btw,
	there are DCL shell's by 3rd parties that at least used to
	be available for ULTRIX)

I've entered QAR 51279 :


 There doesn't seem to be a simple way of causing core files to be
created with unique names.

I'm running parallel applications on a cluster.  A parallel application
(PVM, MPI or HPF) typically consists of multiple copies of the same code
running on one or more nodes of the cluster.  The multiple processes
 often expect to be running in the same working directory.  When they
crash, they all create a file called "core".  By Sods Law, the core file
with the info you're looking for is never the one that remains on the disk
after they've all finished overwriting each other.

I'd like a mechanism, such as setting an environment variable, to indicate
that I'd like unique names for any core files produced.  I'd be satisfied
with appending the pid to the name - better yet would be appending
nodename and pid - better yet would be appending program name,host and pid.

The "High Performance Computing" and "Scientific and Technical" markets
are making more and more use of parallel codes like this, so a need for
such a mechanism is becoming more urgent.

As a workaround, how about implementing a last chance
handler that cooperates w/ other copies of the same
program such that only one process can die at a time,
and they rename old core files before they exit

You'd probably need to find the last-chance handler
first, and call that from your last-chance handler
(or whatever it takes to make the core dump be
written).

I don't know if this would work -- just a thought...

Re: .11

	if by last chance, you mean a signal handler, yes, that could
	work as well.  I've used such signal handlers in the past
	to syslog info from the signal context about the pc and address
	for segv and bus violations.  Such a handler could also rename
	an existing "core" file before re-asserting the signal (and make
	sure to restore the signal's behaviour to SIG_DFL before re-
	asserting).