Thread: file descriptors
We have started getting the error
FATAL: terminating connection due to administrator command
in some of our processes. Searching in the archives, I gather that this is caused by a SIGTERM, and might be coming from a ulimit problem.
We are running Solaris 9/Intel, and the ulimit for nofiles for the database owner process is 256. I suspect this needs to be set to "unlimited", which I don't think should cause a problem on Solaris (?).
Is there any rule of thumb for determining how many file descriptors a given postmaster would use, e.g., a way to find out how many data files exist on the server, and/or an estimate for how many of them would have to be open at any given time? Do ALL of the data files get kept open, or do they get opened and closed?
Thanks.
- DAP
----------------------------------------------------------------------------------
David Parker Tazz Networks (401) 709-5130
"David Parker" <dparker@tazznetworks.com> writes:
> We have started getting the error
> FATAL: terminating connection due to administrator command
> in some of our processes. Searching in the archives, I gather that this
> is caused by a SIGTERM, and might be coming from a ulimit problem.
It is coming from a SIGTERM, but I'm not aware of any platforms that
respond to exceeding the ulimit open-files limit by SIGTERM'ing the
process. I think you're barking up the wrong tree.
> We are running Solaris 9/Intel, and the ulimit for nofiles for the
> database owner process is 256. I suspect this needs to be set to
> "unlimited", which I don't think should cause a problem on Solaris (?).
I think it *would* cause a problem, unless Solaris can support unlimited
numbers of open files --- we have certainly seen PG eat all available
file table slots on other kernels. I don't recommend raising nofiles.
The backends are perfectly capable of working within the nofiles limit
you set, and 256 seems high enough to avoid thrashing.
regards, tom lane
OK, well that's good to know. You mentioned ulimit in http://archives.postgresql.org/pgsql-bugs/2003-12/msg00080.php which ulimit parameters were you thinking of? That post is what set me barking up this tree ;-) The only other thing not set to "unlimited" is stack, which is set to 8480 for the database owner on this system (not sure where that number came from). Thanks. - DAP >-----Original Message----- >From: Tom Lane [mailto:tgl@sss.pgh.pa.us] >Sent: Friday, February 11, 2005 6:17 PM >To: David Parker >Cc: pgsql-general@postgresql.org >Subject: Re: [GENERAL] file descriptors > >"David Parker" <dparker@tazznetworks.com> writes: >> We have started getting the error >> FATAL: terminating connection due to administrator >command in some >> of our processes. Searching in the archives, I gather that this is >> caused by a SIGTERM, and might be coming from a ulimit problem. > >It is coming from a SIGTERM, but I'm not aware of any >platforms that respond to exceeding the ulimit open-files >limit by SIGTERM'ing the process. I think you're barking up >the wrong tree. > >> We are running Solaris 9/Intel, and the ulimit for nofiles for the >> database owner process is 256. I suspect this needs to be set to >> "unlimited", which I don't think should cause a problem on >Solaris (?). > >I think it *would* cause a problem, unless Solaris can support >unlimited numbers of open files --- we have certainly seen PG >eat all available file table slots on other kernels. I don't >recommend raising nofiles. >The backends are perfectly capable of working within the >nofiles limit you set, and 256 seems high enough to avoid thrashing. > > regards, tom lane >
"David Parker" <dparker@tazznetworks.com> writes: > OK, well that's good to know. You mentioned ulimit in > http://archives.postgresql.org/pgsql-bugs/2003-12/msg00080.php Well, that was a different scenario --- or at least I thought it was. Are you seeing unwanted shutdown of the entire database cluster, or just an individual backend? What shows up in the postmaster log when this happens? regards, tom lane
It's just an individual backend. Unfortunately I don't have the postgres log file yet, which would obviously help - I only have the application client's log, but the people reporting the problem know they need to get me the database log file next time it happens. I haven't been able to repro this on my own yet. Thanks. - DAP >-----Original Message----- >From: Tom Lane [mailto:tgl@sss.pgh.pa.us] >Sent: Friday, February 11, 2005 6:40 PM >To: David Parker >Cc: pgsql-general@postgresql.org >Subject: Re: [GENERAL] file descriptors > >"David Parker" <dparker@tazznetworks.com> writes: >> OK, well that's good to know. You mentioned ulimit in >> http://archives.postgresql.org/pgsql-bugs/2003-12/msg00080.php > >Well, that was a different scenario --- or at least I thought it was. >Are you seeing unwanted shutdown of the entire database >cluster, or just an individual backend? What shows up in the >postmaster log when this happens? > > regards, tom lane >