multi-gpu on a node

[Noriyuki Kushida]

Hi,

Thanks to your support, I am (slowly) progressing.

I have just encountered another question, and if you could advise me I am grateful.

Namely, how can I assign GPUs on a node to processes?

I found the following article written by Tabuchi-san, but I need further details.

For example, I am not sure how I can get "acc_device_nvidia".

https://ipsj.ixsq.nii.ac.jp/ej/?action=repository_action_common_download&item_id=102313&item_no=1&attribute_id=1&file_no=1

Alternatively, I may be able to use ACC_SET_DEVICE_NUM.

However, since I am using PGI compiler, I am not sure how I can use the accel_lib module.

Native Compiler: PGI 20.1 Fortran

Omni Compiler: 1.3.2

Best regards,

Noriyuki Kushida

[Masahiro Nakao]

Dear Kushida-san,

Sorry for the late reply.

Tabuchi-san graduated from development, so we do not know the details.
The function xacc_create_device() written in the paper doesn't seem to
be in omni-compiler package.
And omni-compiler cannot use the accel_lib module file.

Instead, you may use the ACC_DEVICE_NUM environment variable, which is
defined in OpenACC specification.
This variable controls the default device number.

If you use OpenMPI included in PGI compiler,
you can execute the following command using the "numa.sh" script,
which is attached in this mail.

$ mpirun -np 2 sh ./numa.sh ./a.out

Thanks,

> --
> You received this message because you are subscribed to the Google
> Groups "omni-compiler" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to omni-compile...@googlegroups.com
> <mailto:omni-compile...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/omni-compiler/CAKx-aKysdLUJwf%3Dv%2B2vMYeWp-oHPe3Lw6u%2BtoTSOheZcVL81gA%40mail.gmail.com
> <https://groups.google.com/d/msgid/omni-compiler/CAKx-aKysdLUJwf%3Dv%2B2vMYeWp-oHPe3Lw6u%2BtoTSOheZcVL81gA%40mail.gmail.com?utm_medium=email&utm_source=footer>.

--
Masahiro NAKAO
Email : masahir...@riken.jp
Researcher
RIKEN Center for Computational Science (R-CCS)