[CP2K-user] [CP2K:15203] Re: Does CP2K allow a multi-GPU run?

Lenard Carroll lenardc... at gmail.com
Fri Apr 23 07:59:53 UTC 2021
Previous message (by thread): [CP2K-user] [CP2K:15203] Re: Does CP2K allow a multi-GPU run?
Next message (by thread): [CP2K-user] [CP2K:15203] Re: Does CP2K allow a multi-GPU run?
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Dear Fabian. COSMA wasn't installed with CP2K, so that can't be the issue.
The HPC system is not CRAY, but I did ask the HPC admin to look into it.

On Thu, Apr 22, 2021 at 8:02 PM fa... at gmail.com <fabia... at gmail.com>
wrote:

> Hi,
>
> cp2k is crashing when COSMA tries to access a gpu ("error: GPU API call :
> invalid resource handle"). On cray systems there is the environment
> variable "export CRAY_CUDA_MPS=1" that has to be set. Otherwise only one
> mpi rank can access a specific GPU device. Maybe there is a similar setting
> for your cluster?
>
> Also cp2k can be memory hungry. Setting "ulimit -s unlimited" is often
> needed.
>
> I hope this helps,
> Fabian
>
> On Thursday, 22 April 2021 at 19:36:35 UTC+2 ASSIDUO Network wrote:
>
>> Oh you meant the error file. Please find it attached.
>>
>> I have run on CPU only and one GPU. It works.
>>
>> On Thu, Apr 22, 2021 at 7:31 PM Alfio Lazzaro <al... at gmail.com>
>> wrote:
>>
>>> I'm sorry, I cannot assist you, I'm not an expert on how to use CP2K ('m
>>> not a domain scientist). Without the total log, I can help you...
>>> I assume you should have a log file from PBS where you can see the error
>>> message. I can assume it is a memory limit.
>>> Have you executed on a CPU only?
>>>
>>>
>>>
>>> Il giorno giovedì 22 aprile 2021 alle 17:45:06 UTC+2 ASSIDUO Network ha
>>> scritto:
>>>
>>>> Here's the log file. The job ended prematurely.
>>>>
>>>> On Thu, Apr 22, 2021 at 3:23 PM Lenard Carroll <len... at gmail.com>
>>>> wrote:
>>>>
>>>>> Not sure yet. The job is still in the queue. As soon as it is finished
>>>>> I'll post the log file info here.
>>>>>
>>>>> On Thu, Apr 22, 2021 at 3:15 PM Alfio Lazzaro <al... at gmail.com>
>>>>> wrote:
>>>>>
>>>>>> And it works? Check the output and the performance... It can be that
>>>>>> your particular test case doesn't use the GPU at all, so could you attach
>>>>>> the log (at least the final part of it)
>>>>>>
>>>>>> Il giorno giovedì 22 aprile 2021 alle 13:42:16 UTC+2 ASSIDUO Network
>>>>>> ha scritto:
>>>>>>
>>>>>>> I am using 30 threads now over 3 GPUs, so I used:
>>>>>>>
>>>>>>> export OMP_NUM_THREADS=10
>>>>>>> mpiexec -n 3 cp2k.psmp -i gold50.inp -o gold50.out
>>>>>>>
>>>>>>>
>>>>>>> On Thu, Apr 22, 2021 at 1:34 PM Alfio Lazzaro <al... at gmail.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Wait, I see you have 32 threads in total, so need to have 32/4 = 8
>>>>>>>> threads.
>>>>>>>> Please change
>>>>>>>>
>>>>>>>> export OMP_NUM_THREADS=8
>>>>>>>>
>>>>>>>> Il giorno giovedì 22 aprile 2021 alle 13:27:59 UTC+2 ASSIDUO
>>>>>>>> Network ha scritto:
>>>>>>>>
>>>>>>>>> Shall do. I already set it up, but it's in a long queue.
>>>>>>>>>
>>>>>>>>> On Thu, Apr 22, 2021 at 1:22 PM Alfio Lazzaro <al... at gmail.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> Could you try what I suggested:
>>>>>>>>>>
>>>>>>>>>> export OMP_NUM_THREADS=10
>>>>>>>>>> mpirun -np 4 ./cp2k.psmp -i gold.inp -o gold_pbc.out
>>>>>>>>>>
>>>>>>>>>> Please check the corresponding log.
>>>>>>>>>>
>>>>>>>>>> As I said above, you need an MPI rank per GPU and you told us
>>>>>>>>>> that you have 4 GPUs, so you need 4 ranks (or multiple). With 10 you get
>>>>>>>>>> unbalance.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Il giorno giovedì 22 aprile 2021 alle 10:17:27 UTC+2 ASSIDUO
>>>>>>>>>> Network ha scritto:
>>>>>>>>>>
>>>>>>>>>>> Correction, he told me to use:
>>>>>>>>>>>
>>>>>>>>>>> mpirun -np 10 cp2k.psmp -i gold.inp -o gold_pbc.out
>>>>>>>>>>>
>>>>>>>>>>> but it didn't run correctly.
>>>>>>>>>>>
>>>>>>>>>>> On Thu, Apr 22, 2021 at 9:51 AM Lenard Carroll <
>>>>>>>>>>> len... at gmail.com> wrote:
>>>>>>>>>>>
>>>>>>>>>>>> He suggested I try out:
>>>>>>>>>>>> mpirun -n 10 cp2k.psmp -i gold.inp -o gold_pbc.out
>>>>>>>>>>>>
>>>>>>>>>>>> as he is hoping that will cause the 1 GPU to use 10 CPUs over
>>>>>>>>>>>> the selected 4 GPUs.
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> On Thu, Apr 22, 2021 at 9:48 AM Alfio Lazzaro <
>>>>>>>>>>>> al... at gmail.com> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>> Your command to run CP2K doesn't mention MPI (mpirun, mpiexc,
>>>>>>>>>>>>> ...). Are you running with multiple ranks?
>>>>>>>>>>>>>
>>>>>>>>>>>>> You can check those lines in the output:
>>>>>>>>>>>>>
>>>>>>>>>>>>>  GLOBAL| Total number of message passing processes
>>>>>>>>>>>>>                 32
>>>>>>>>>>>>>  GLOBAL| Number of threads for this process
>>>>>>>>>>>>>                 4
>>>>>>>>>>>>>
>>>>>>>>>>>>> And check your numbers.
>>>>>>>>>>>>> I can guess you have 1 rank and 40 threads.
>>>>>>>>>>>>> To use 4 GPUs you need 4 ranks (and less threads per rank),
>>>>>>>>>>>>> i.e. something like
>>>>>>>>>>>>>
>>>>>>>>>>>>> export OMP_NUM_THREADS=10
>>>>>>>>>>>>> mpiexec -n 4 ./cp2k.psmp -i gold.inp -o gold_pbc.out
>>>>>>>>>>>>>
>>>>>>>>>>>>> Please check with your sysadmin on how to run with multiple
>>>>>>>>>>>>> MPI ranks.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Hope it helps.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Alfio
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> Il giorno mercoledì 21 aprile 2021 alle 09:26:53 UTC+2 ASSIDUO
>>>>>>>>>>>>> Network ha scritto:
>>>>>>>>>>>>>
>>>>>>>>>>>>>> This is what my PBS file looks like:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> #!/bin/bash
>>>>>>>>>>>>>> #PBS -P <PROJECT>
>>>>>>>>>>>>>> #PBS -N <JOBNAME>
>>>>>>>>>>>>>> #PBS -l select=1:ncpus=40:ngpus=4
>>>>>>>>>>>>>> #PBS -l walltime=08:00:00
>>>>>>>>>>>>>> #PBS -q gpu_4
>>>>>>>>>>>>>> #PBS -m be
>>>>>>>>>>>>>> #PBS -M none
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> module purge
>>>>>>>>>>>>>> module load chpc/cp2k/8.1.0/cuda10.1/openmpi-4.0.0/gcc-7.3.0
>>>>>>>>>>>>>> source $SETUP
>>>>>>>>>>>>>> cd $PBS_O_WORKDIR
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> cp2k.psmp -i gold.inp -o gold_pbc.out
>>>>>>>>>>>>>> ~
>>>>>>>>>>>>>>                                                            ~
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> On Wed, Apr 21, 2021 at 9:22 AM Alfio Lazzaro <
>>>>>>>>>>>>>> al... at gmail.com> wrote:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> The way to use 4 GPUs per node is to use 4 MPI ranks. How
>>>>>>>>>>>>>>> many ranks are you using?
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Il giorno martedì 20 aprile 2021 alle 19:44:15 UTC+2 ASSIDUO
>>>>>>>>>>>>>>> Network ha scritto:
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> I'm asking, since the administrator running my country's
>>>>>>>>>>>>>>>> HPC is saying that although I'm requesting access to 4 GPUs, CP2K is only
>>>>>>>>>>>>>>>> using 1. I checked the following output:
>>>>>>>>>>>>>>>>  DBCSR| ACC: Number of devices/node
>>>>>>>>>>>>>>>>                     4
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> And it shows that CP2K is picking up 4 GPUs.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> On Tuesday, April 20, 2021 at 3:00:17 PM UTC+2 ASSIDUO
>>>>>>>>>>>>>>>> Network wrote:
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> I currently have access to 4 GPUs to run an AIMD
>>>>>>>>>>>>>>>>> simulation, but only one of the GPUs are being used. Is there a way to use
>>>>>>>>>>>>>>>>> the other 3, and if so, can you tell me how to set it up with a PBS job?
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> --
>>>>>>>>>>>>>>> You received this message because you are subscribed to the
>>>>>>>>>>>>>>> Google Groups "cp2k" group.
>>>>>>>>>>>>>>> To unsubscribe from this group and stop receiving emails
>>>>>>>>>>>>>>> from it, send an email to cp... at googlegroups.com.
>>>>>>>>>>>>>>> To view this discussion on the web visit
>>>>>>>>>>>>>>> https://groups.google.com/d/msgid/cp2k/70ba0fce-8636-4b75-940d-133ce4dbf0can%40googlegroups.com
>>>>>>>>>>>>>>> <https://groups.google.com/d/msgid/cp2k/70ba0fce-8636-4b75-940d-133ce4dbf0can%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>>>>>>>>>>>> .
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>> --
>>>>>>>>>>>>> You received this message because you are subscribed to the
>>>>>>>>>>>>> Google Groups "cp2k" group.
>>>>>>>>>>>>> To unsubscribe from this group and stop receiving emails from
>>>>>>>>>>>>> it, send an email to cp... at googlegroups.com.
>>>>>>>>>>>>> To view this discussion on the web visit
>>>>>>>>>>>>> https://groups.google.com/d/msgid/cp2k/92e4f88d-fde8-4127-ab5f-0b98bbbba8ebn%40googlegroups.com
>>>>>>>>>>>>> <https://groups.google.com/d/msgid/cp2k/92e4f88d-fde8-4127-ab5f-0b98bbbba8ebn%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>>>>>>>>>> .
>>>>>>>>>>>>>
>>>>>>>>>>>> --
>>>>>>>>>> You received this message because you are subscribed to the
>>>>>>>>>> Google Groups "cp2k" group.
>>>>>>>>>> To unsubscribe from this group and stop receiving emails from it,
>>>>>>>>>> send an email to cp... at googlegroups.com.
>>>>>>>>>>
>>>>>>>>> To view this discussion on the web visit
>>>>>>>>>> https://groups.google.com/d/msgid/cp2k/59a635d8-0f0c-4dc5-abaf-b8bbe3c18da5n%40googlegroups.com
>>>>>>>>>> <https://groups.google.com/d/msgid/cp2k/59a635d8-0f0c-4dc5-abaf-b8bbe3c18da5n%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>>>>>>> .
>>>>>>>>>>
>>>>>>>>> --
>>>>>>>> You received this message because you are subscribed to the Google
>>>>>>>> Groups "cp2k" group.
>>>>>>>> To unsubscribe from this group and stop receiving emails from it,
>>>>>>>> send an email to cp... at googlegroups.com.
>>>>>>>>
>>>>>>> To view this discussion on the web visit
>>>>>>>> https://groups.google.com/d/msgid/cp2k/ec4efd81-6314-4ce7-b22c-148b362d2ba6n%40googlegroups.com
>>>>>>>> <https://groups.google.com/d/msgid/cp2k/ec4efd81-6314-4ce7-b22c-148b362d2ba6n%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>>>>> .
>>>>>>>>
>>>>>>> --
>>>>>> You received this message because you are subscribed to the Google
>>>>>> Groups "cp2k" group.
>>>>>> To unsubscribe from this group and stop receiving emails from it,
>>>>>> send an email to cp... at googlegroups.com.
>>>>>> To view this discussion on the web visit
>>>>>> https://groups.google.com/d/msgid/cp2k/d29306aa-e0b8-4797-9298-13dab23e9083n%40googlegroups.com
>>>>>> <https://groups.google.com/d/msgid/cp2k/d29306aa-e0b8-4797-9298-13dab23e9083n%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>>> .
>>>>>>
>>>>> --
>>> You received this message because you are subscribed to the Google
>>> Groups "cp2k" group.
>>> To unsubscribe from this group and stop receiving emails from it, send
>>> an email to cp... at googlegroups.com.
>>>
>> To view this discussion on the web visit
>>> https://groups.google.com/d/msgid/cp2k/6852eb71-6886-4fe7-8f4a-3ad8318a289dn%40googlegroups.com
>>> <https://groups.google.com/d/msgid/cp2k/6852eb71-6886-4fe7-8f4a-3ad8318a289dn%40googlegroups.com?utm_medium=email&utm_source=footer>
>>> .
>>>
>> --
> You received this message because you are subscribed to the Google Groups
> "cp2k" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to cp... at googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/cp2k/c2033277-5fdf-4e98-9329-e9a289a5b277n%40googlegroups.com
> <https://groups.google.com/d/msgid/cp2k/c2033277-5fdf-4e98-9329-e9a289a5b277n%40googlegroups.com?utm_medium=email&utm_source=footer>
> .
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20210423/2d031eff/attachment.htm>
Previous message (by thread): [CP2K-user] [CP2K:15203] Re: Does CP2K allow a multi-GPU run?
Next message (by thread): [CP2K-user] [CP2K:15203] Re: Does CP2K allow a multi-GPU run?
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the CP2K-user mailing list