Thread: Postgres process resident size does not drop after killing statement

Every so often our production Postgres 8.3 system will get statement
that runs for a few hours, or a few days, or more, and needs to be
killed dead.  We kill it with pg_cancel_backend(), and cpu usage of the
process immediately drops, and the process starts serving other
statements.  But the curious thing is the resident size does not drop. 
Is this normal?  Something to be concerned about?  Here's a mild example
(only 2.8 gigs):

# top PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
20108 postgres  16   0 4305m 2.8g 2.8g D    4  8.6   3:12.50 postgres:
production production 10.100.4.11(44200) SELECT
16306 postgres  15   0 4248m 1.5g 1.5g S    0  4.7   0:47.70 postgres:
jira jira 10.100.5.12(57517) idle
28472 postgres  15   0 4249m 1.4g 1.4g S    2  4.6   0:16.18 postgres:
production production 10.100.4.11(48457) idle

# uname -a
Linux xxxx 2.6.18-6-amd64 #1 SMP Thu Dec 25 21:08:54 UTC 2008 x86_64
GNU/Linux

Bryce Nesbitt <bryce2@obviously.com> writes:
> Every so often our production Postgres 8.3 system will get statement
> that runs for a few hours, or a few days, or more, and needs to be
> killed dead.  We kill it with pg_cancel_backend(), and cpu usage of the
> process immediately drops, and the process starts serving other
> statements.  But the curious thing is the resident size does not drop. 
> Is this normal?

Resident size as displayed by "top" is pretty misleading on many
systems.  Typically, it counts whatever private memory the process
has plus as many pages of the shared memory area as the process has
actually touched in its lifespan.  So the latter portion ramps up
from nil to approach the size of the shared memory block just as a
byproduct of normal operations.  You did not say what your
shared_buffers setting is, but if it's a GB or two then I don't
find anything surprising in the numbers you show.
        regards, tom lane