Re: total memory of children processes

[Martin Pala]

I see the same problem on my station too:

unicorn:~/cvs/monit# ps -leaf
F S UID        PID  PPID  C PRI  NI ADDR SZ WCHAN  STIME TTY          
TIME CMD
1 S root      4890     1  0  69   0 -  2604 rt_sig 18:23 ?        
00:00:00 /usr/sbin/slapd
1 S root      4891  4890  0  69   0 -  2604 poll   18:23 ?        
00:00:00 /usr/sbin/slapd
1 S root      4892  4891  0  69   0 -  2604 select 18:23 ?        
00:00:00 /usr/sbin/slapd
1 S root      4913  4891  0  69   0 -  2604 rt_sig 18:24 ?        
00:00:00 /usr/sbin/slapd

unicorn:~/cvs/monit# ./monit -c /etc/monitrc status
Process 'slapd' is running with pid [4890]
       Uptime: 3m  CPU: 0.0% Memory w/o children: 0.8% [2604kB]
       Children: 3 Memory w/ children: 3.2% [10416kB]
       Monitoring status: monitored

As you can see, the reported size is sum of RSS off all threads 
(4*2604=10416). This is not correct result, because 2604kB is memory 
shared by threads (LinuxThreads reports each thread with its PID and 
size, but the memory is shared).


Note: this is not related to /proc/kcore patch - it only replaced method 
for available memory detection (instead of kcore size it is read from 
/proc/meminfo - the result is the same) => this is standalone problem (bug).

Martin


Arkadiusz Miskiewicz wrote:

> Hi,
>
> Total memory of children processes seems miscalculated:
> Event: total mem amount of 12193680kB matches resource limit [total mem 
> amount>512001kB]
>
> ok but this machine has only 512MB ram and 2GB of swap.
>
> The problem seems be that only rss (resident set size) value is used. rss also 
> includes memory shared between processes.
>
> IMO algoritm to calculate total mem amount should looke like:
>
> rss of parent + (rss of first child - shared mem, rss of second chold - shared 
> mem and so on). It's still not perfect because we don't know between which 
> processes that memory is shared but it's much better than adding whole rss.
>
> shared memory can be found in /proc/$pid/statm.
>
> On my server for httpd proces it's:
> address@hidden root]# pmap -x 22746 |less
> 22746:   httpd -4
> Address       kB Resident Shared Private Permissions       Name
> 00110000      80       -       0      80 read/exec         ld-2.2.5.so
> 00124000       4       -       0       4 read/write        ld-2.2.5.so
> 00125000       8       -       0       8 read/write         [ anon ]
> 00128000       4       -       0       4 read/exec         mod_asis.so
> [...]
> 01142000     980       -     980       0 read/write        mod_jk2.shm
> 01237000     980       -     980       0 read/write        mod_jk2.shm
> 0132c000     508       -     508       0 read/write        SYSV00000000 
> (deleted)
> 013ab000     364       -     364       0 read/write        SYSV00000000 
> (deleted)
> 08048000     212       -       0     212 read/exec         httpd
> 0807d000      32       -       0      32 read/write        httpd
> 08085000   63820       -       0   63820 read/write/exec    [ anon ]
> 40000000   32768       -   32768       0 read/write        SYSV00000000 
> (deleted)
> 42000000   32768       -   32768       0 read/write        SYSV00000000 
> (deleted)
> bfff5000      44       -       0      44 read/write/exec    [ anon ]
> --------  ------  ------  ------  ------
> total kB  148912       -   71360   77552
>
> so almost half of memory used by this process is shared between httpd (and 
> other) processes...
>
> Comments?
>