On Sun, Mar 20, 2011 at 05:22:33PM -0400, Dave Love wrote:
Jesse Becker <[email protected]> writes:
Are you having a problem that the limits are not set by SGE, or that they
are not acutally enforced by the OS? I've had a lot of trouble on Linux
systems where the "data" and "rss" settings are ignored. The "as"
limit, however, is enforced.
Rss is documented as ignored in recent Linuxen -- see setrlimit(2). I
doubt data is ignored, but it's not what you normally want, e.g. it
doesn't limit memory malloced via mmap. (I speak from experience of
setting the wrong limits.)
It was ignored in my testing. I have a simple memory allocation
program, and set rss, data, and as to various different limits so as to
watch when the malloc() failed. The testing was, in essence:
1) set the limilts rss=10M, data=20M, as=30M,
2) allocate memory in 1M chunks and see where it fails
The malloc failed at 30M.
I then set the limits: as=10M, rss=20M, data=30M, and ran the program
again.
The malloc failed at 10M.
Thus, for my *simple* testing, data and rss are both ignored (and yes,
rss *is* documented as being ignored, but data is not).
You normally want to limit the address space (a hard "as" limit is
equivalent to GE's h_vmem). It's unfortunate that you can't exclude
non-anonymous mmap from that, though, which has come up before.
--
Jesse Becker
NHGRI Linux support (Digicon Contractor)
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
