Tom Lane presumably uttered the following on 06/29/05 19:12:
Sven Willenberger <sven@xxxxxxx> writes:
I have found the answer/problem. On a hunch I increased maxdsiz to 1.5G
in the loader.conf file and rebooted. I ran vacuumdb and watched top as
the process proceeded. What I saw was SIZE sitting at 603MB (which was
512MB plus another 91MB which corresponded nicely to the value of RES
for the process. A bit into the process I saw SIZE jump to 1115 -- i.e.
another 512 MB of RAM was requested and this time allocated. At one
point SIZE dropped back to 603 and then back up to 1115. I suspect the
same type of issue was occuring in regular vacuum from the psql client
connecting to the backend, for some reason not as frequently. I am
gathering that maintenance work mem is either not being recognized as
having already been allocated and another malloc is made or the process
is thinking the memory was released and tried to grab a chunk of memory
again.
Hmm. It's probably a fragmentation issue. VACUUM will allocate a
maintenance work mem-sized chunk during command startup, but that's
likely not all that gets allocated, and if any stuff allocated after
it is not freed at the same time, the process size won't go back down.
Which wouldn't be a killer in itself, but unless the next iteration
is able to fit that array in the same space, you'd see the above
behavior.
So maintenance work mem is not a measure of the max that can allocated
by a maintenance procedure but rather an increment of memory that is
requested by a maintenance process (which currently are vacuum and
index, no?), if my reading of the above is correct.
BTW, do you have any evidence that it's actually useful to set
maintenance work mem that high for VACUUM? A quick and dirty solution
would be to bound the dead-tuples array size at something more sane...
I was under the assumption that on systems with RAM to spare, it was
beneficial to set main work mem high to make those processes more
efficient. Again my thinking was that the value you set for that
variable determined a *max* allocation by any given maintenance process,
not a memory allocation request size. If, as my tests would indicate,
the process can request and receive more memory than specified in
maintenance work mem, then to play it safe I imagine I could drop that
value to 256MB or so.
Sven
---------------------------(end of broadcast)---------------------------
TIP 6: Have you searched our list archives?
http://archives.postgresql.org
