Inconsistencies in the way process memory is being reported in /proc
From: Mark Seger
Date: Wed Jun 25 2008 - 15:30:35 EST
I have seen past postings on this topic and there still seems to be
issues. Since adding process i/o stats to collectl I've been looking
more closely at process stats in general and have been digging deeper
into how memory is being accounted for on my current test system which
is running a 2.6.23 kernel on top of a rhel4.2 distribution and have
been using 'man proc' as a guide.
Specifically, I realize stats are reported in /proc/pid/stat,
/proc/pid/statm and /proc/pid/status and according to the manpage
'status' is mainly a summary of many of the data from the other 2 in
more human readable form. That said I wrote a little script that grabs
all these structures as once and reports them in a way that lets me
compare them all along with 'ps' outptu and when I do this I start
getting dizzy. For example, when I do a ps I see:
USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND
root 29216 0.0 0.3 81836 15160 ? Ss 11:11 0:09
/usr/bin/perl -w /usr/sbin/collectl -D
statm contains, after multiplying by 4 since these are in pages:
total: 81836
resident: 15160
sharedpages: 1280
text: 16
data/stack: 0
library: 14292
dirty: 0
and status contains:
VmPeak: 81836 kB
VmSize: 81836 kB
VmLck: 0 kB
VmHWM: 15160 kB
VmRSS: 15160 kB
VmData: 14208 kB
VmStk: 84 kB
VmExe: 16 kB
VmLib: 3620 kB
VmPTE: 108 kB
I only saw VSS and RSS in stat and happily they agree with all the other
2 structures as well as ps
However, that's where the good news stops. For example, statm says the
data/stack size is 0 but status shows them as non-zero. Furthermore, it
looks like if I add the data and stacksize from 'status' I get the
'library' size in statm. The library size in 'status' is 3260 and I
don't see any way to correlate that with anything! I also don't see
mention of some of the other fields reorted in 'status'. I'm guessing
peak/hwm are both highwater marks for vss and rss, but what about
VmLck? Can that be derived from anything in 'stat'? I'm also assuming
the size of the exe is constant and probably not particularly useful.
Can anyone shed any light on what numbers to believe or should I only
count on vss/rss being correct? If someone can provide more detailed
definitions I'll be happy to add them to collectl's documentation.
And speaking of collectl, when it run as a daemon it continuously
records all these files (and I'm hoping I can stop grabbing 'status' if
someone can tell me how to derive these numbers from the other 2 data
structures in a believable form) and a number of ways to later display
the data. In one form, it simply reports a bunch of fields from
'status' as shown below:
# PID User S VmSize VmLck VmRSS VmData VmStk VmExe VmLib MajF
MinF Command
3203 root S 19940K 0 436K 416K 84K 48K 3800K
0 0 rpc.idmapd
3291 root S 2876K 0 360K 240K 84K 208K 1276K
0 0 /usr/sbin/smartd
3395 root S 22016K 0 1284K 444K 84K 324K 3496K
0 0 /usr/sbin/sshd
3410 root S 8792K 0 792K 376K 84K 152K 1972K
0 0 xinetd
3420 root S 4260K 0 336K 188K 84K 84K 1808K
0 0 gpm
3435 root S 57132K 0 968K 436K 520K 40K 1492K
0 0 crond
but I also admit these aren't necessarily the most useful fields to
report, especially now that I've been digging deeper. So my second
question is if anyone cares to recommend some alternative fields to
report under the category of 'process memory utilitization', I'd be
happy to consider changing what I'm reporting. Of course all that
assume there is something more interesting than just vss and rss which
I'm sure there must be.
sorry for being so long winded...
-mark
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/