[yt-users] Parallel data reading and covering grid

[Yingchao Lu]

Hi Nathan,

I opened an issue.

Thanks,
Yingchao

On Wed, Sep 20, 2017 at 10:20 AM, Nathan Goldbaum <nathan12343 at gmail.com>
wrote:

> Hi Yingchao,
>
> I don't think covering grid construction is parallelized at the moment -
> in your script each processor is creating its own copy of the covering
> grid. I haven't tried experimenting with this though so it's possible I'm
> misreading the code or missing something happening elsewhere to make this
> operation parallel.
>
> One way around this with the current version of yt would be to handle the
> parallelism yourself - i.e. have each processor construct a subvolume of
> the covering grid and then combine it yourself using e.g. mpi4py.
>
> I think it might also be possible to make yt construct the covering grid
> in a parallel-aware fashion. The relevant code is here:
>
> https://github.com/yt-project/yt/blob/master/yt/data_objects
> /construction_data_containers.py#L686
>
> One would need to make the loop over io chunks a parallel loop and then
> add a reduction step at the end.
>
> If you don't feel like taking on this task yourself please feel free to
> open an issue about it on github so that we don't lose track of the feature
> request.
>
> -Nathan
>
> On Tue, Sep 19, 2017 at 7:53 PM, Yingchao Lu <yingchao.lu at gmail.com>
> wrote:
>
>> Dear yt users,
>>
>> I am trying to read some AMR data and convert them into 3D array. The
>> test code:
>> ######################### BEGIN #########################
>> import yt
>> from time import time
>> yt.enable_parallelism()
>>
>> ds = yt.load("flash_hdf5_plt_cnt_0000")
>> tstart = time()
>> cg = ds.covering_grid(level=ds.max_level, left_edge=ds.domain_left_edge,
>> dims=ds.domain_dimensions*2**ds.max_level)
>> cg['dens']
>>
>> if yt_isroot(): print "It takes {0}s".format(time()-tstart)
>> ######################### END #########################
>>
>> I tried to run to serial or parallel on an interactive session on
>> stampede:
>>  ######################### BEGIN #########################
>> [yclu@ test]$ ls
>> flash_hdf5_plt_cnt_0000  test.py
>> [yclu@ test]$ python test.py
>> It takes 34.0571820736s
>> [yclu@ test]$ export OMP_NUM_THREADS=68
>> [yclu@ test]$ python test.py
>> It takes 33.1969199181s
>> [yclu@ test]$ export OMP_NUM_THREADS=1
>> [yclu@ test]$ mpirun -np 68 python test.py
>> It takes 58.0391800404s
>> ######################### END #########################
>>
>> The time does not seem to be reduced by parallelism. And multi-process
>> seem to have huge communication overhead. Is there a way to increase the
>> speed by parallelism?
>>
>> Thanks,
>> Yingchao
>>
>>
>> _______________________________________________
>> yt-users mailing list
>> yt-users at lists.spacepope.org
>> http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
>>
>>
>
> _______________________________________________
> yt-users mailing list
> yt-users at lists.spacepope.org
> http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.spacepope.org/pipermail/yt-users-spacepope.org/attachments/20170920/8a853efc/attachment.html>