[Pw_forum] Memory usage by pw.x

[Axel Kohlmeyer]

guido,

On Thu, Aug 30, 2012 at 12:04 PM, Guido Fratesi <fratesi at mater.unimib.it> wrote:
> I'm sorry for a previous incomplete message.
>
>>   sed -ri "s/(^ *)(allocate.*$)/\1\2\n\1  CALL mem_whatever()/i" $(find
>> /where/is/espresso -name \*.f90)
>
> That was very useful, thank you.
>
> I now get 1.9GB out of 2.4, which starts giving some usable estimate,
> but I do understand that getting the accurate value is very complex.

quantifying memory usage precisely on unix/linux machines with
virtual memory management and memory sharing is almost
impossible. you have multiple components to worry about:

- address space (memory reserved to be used, but initially all
  mapped to the same copy-on-write location)
- resident set size (actual physical memory used)
- shared memory (it is not memory that is shared, but
   more a measure for how much sharing is going on)
- swap space.
- device memory (from infiniband cards for example)
- pinned memory (allocated memory that cannot be swapped,
   usually used to back device memory)

so what is the real memory usage is difficult to determine.
address space (VMEM) is usually too large, resident set size
(RSS) does not consider memory that is swapped out, so it
is often too small. using many MPI tasks drives up the address
space for device memory (which doesn't increase real memory
usage, but also requires more pinned memory, which makes
swapping more likely). multi-threading results in a lot of sharing.

...and tracking allocations in the code only handles explicit
allocations, not those incurred by the fortran language on the
stack or otherwise.

so for all practical purposes you can say that memory use is
usually somewhere between VMEM and RSS, but that can be
pretty far apart.

axel.

> To my purpose, I'll monitor the memory occupancy by a script, which
> follows below in case someone finds it useful...



> Guido
>
> ================
>
> command=pw.x
>
> maxsecs=$((60*60*24))
> delay=1
> nsteps=$((maxsecs/delay))
>
> echo "#when  RAM  PID (0 for all $command instances)"}
>
> for ((i=0;((i<nsteps));i++)); do
>    #timer=`date +%H:%M:%S`
>    #timer=`date +"%H:%M:%S %s.%N"`
>    timer=`date +"%s.%N"`
>    ps -eo comm,rss,pid |
>      awk -v comm=$command -v timer=$timer '
>        BEGIN {tot=0}
>        ($1==comm) {
>          tot+=$2;
>          print timer, $2, $3;
>        }
>        END {if (tot) print timer, tot, 0}
>      '
>    sleep 1
> done
>
>
>
> --
> Guido Fratesi
>
> Dipartimento di Scienza dei Materiali
> Universita` degli Studi di Milano-Bicocca
> via Cozzi 53, 20125 Milano, Italy
>
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum



-- 
Dr. Axel Kohlmeyer  akohlmey at gmail.com  http://goo.gl/1wk0
International Centre for Theoretical Physics, Trieste. Italy.