[ncl-talk] Is there a way to run NCL on multiple processors?
Dennis Shea
shea at ucar.edu
Mon Oct 13 13:40:06 MDT 2014
This is often called 'task prallelism' or 'embarrasingly parallel'.
>From the WWW: "A problem that is obviously decomposable into many identical
but separate subtasks is called 'embarrasingly parallel.'
==
There is tool developed within our section called the CVDP (Climate
Variability Diagnostics Processor). NCL is used to perform shell scripting
type tasks and to do the diagnostics which include EOFs, assorted indices,
etc, netCDF creation and graphics for 100+ ensemble members. It can take
3+days to run in serial mode. Each ensemble run is independent of all other
ensemble runs.
I talked with a software engineer within NCAR's computing division and she
proposed a python approach. She called it 'python light'. Here is a crude
prototype of what it might look like. It uses NCL's 'command line
arguments'.
=====
#!/usr/bin/env python
import parUtils as par
from mpi4py import MPI
import os
# the following is an ascii (text) file which contains the names of
# ensembles members to process and the periods
namelist_fn = 'namelist.CESMcomparison'
# open the namelist file and store it in a file pointer variable
# think of this as NCL's asciiread. Each line is a different ensemble member
namelist_fp = open(namelist_fn,"r")
# read the file and split the lines into an array
# think of this as NCL's str_get_field
nl_lines = namelist_fp.read().splitlines()
# partition the lines up between the mpi tasks
local_nl_lines = par.Partition(nl_lines)
# for each line in the orig nl file that this task is responsible for,
# parse the line and call ncl with the variables
for line in local_nl_lines:
#split line at "|"
name,location,start,stop = line.split("|")
os.system("ncl 'name=\""+name+"\"' 'location=\""+location+"\"'
'start=\""+start+"\"' 'stop=\""+stop+"\"' CVDP.ncl")
# Wait for all procs to get to this point
par.Sync()
=======
The advantage is that python does the book keeping.
=======
Nothing much has gone beyond the above. I am just too busy to focus on this
right now.
D
On Mon, Oct 13, 2014 at 12:29 PM, jbuzan <jbuzan at purdue.edu> wrote:
> Yes, I agree. I recommend using shell scripts that feed NCL the
> specifics. For example, I have scripts that will submit 100s of ncl
> routines that are executed on individual processors on a cluster. This
> aides you in a couple of ways: 1) you have short NCL scripts, which reduce
> programming time, 2) they are less likely to have errors script errors, and
> 3) you speed up your throughput dramatically.
>
> -Jonathan
>
>
> On Oct 13, 2014, at 2:22 PM, Colin Zarzycki <zarzycki at umich.edu> wrote:
>
> It's worth pointing out that for some applications, you can write shell
> scripts which can pseudo-parallelize NCL by running multiple instances
> (which are in serial) alongside one another.
>
> Ex: If you have a high-resolution file which contains 20 time records and
> you would like to use NCL to calculate vorticity from U/V for each of the
> 20 time records, you could spawn 20 instances of NCL which each operate on
> 1 of the time records. Then concatenate the output in serial after each of
> the 20 jobs finishes. This will be much faster than reading the entire
> dataset into a single instance of NCL.
>
> I have found this technique highly useful when processing very large
> datasets (or large numbers of files).
>
> -Colin
>
> On Sun, Oct 12, 2014 at 9:29 PM, 杨玉华 <yangyh at mail.typhoon.gov.cn> wrote:
>
>> Dear Kyle Griffin,
>>
>> Thanks for your great information.
>>
>> Yuhua
>>
>>
>> *您于 Sun, 12 Oct 2014 22:03:49 -0500 的来信中写到: *
>>
>> There is the ParVis project, based on being able to run a select few
>> (but crucial) functions from NCL on multiple processors:
>>
>> http://trac.mcs.anl.gov/projects/parvis/wiki
>>
>> Unfortunately, the only binaries provided are for system types not widely
>> used in the atmospheric science community (as far as I'm aware) and the
>> build instructions aren't particularly detailed. As such, I've never gotten
>> the ncl included in the ParVis project to work. There is a separate mailing
>> list for the discussion, should you have any problems and/or success with
>> the software. Beyond this, though, I'm not aware of any other efforts to
>> parallelize NCL.
>>
>>
>> Kyle
>>
>> ----------------------------------------
>> Kyle S. Griffin
>> Department of Atmospheric and Oceanic Sciences
>> University of Wisconsin - Madison
>> Room 1421
>> 1225 W Dayton St, Madison, WI 53706
>> Email: ksgriffin2 at wisc.edu
>>
>> 2014-10-12 21:56 GMT-05:00 杨玉华 <yangyh at mail.typhoon.gov.cn>:
>>
>>> Hi all,
>>> Is there a way to run NCL on multiple processors?
>>> Thanks,
>>> Yuhua
>>>
>>> ＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝
>>> 名称： 中国气象局上海台风研究所
>>> 地址： 中国上海市徐汇区蒲西路166号
>>> 电话：（＋86）021-64391966
>>> 传真：（＋86）021-64391966
>>> ============================
>>>
>>> _______________________________________________
>>> ncl-talk mailing list
>>> List instructions, subscriber options, unsubscribe:
>>> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>>>
>>>
>>
>> ＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝＝
>> 名称： 中国气象局上海台风研究所
>> 地址： 中国上海市徐汇区蒲西路166号
>> 电话：（＋86）021-64391966
>> 传真：（＋86）021-64391966
>> ============================
>>
>> _______________________________________________
>> ncl-talk mailing list
>> List instructions, subscriber options, unsubscribe:
>> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>>
>> --
>> <http://mailman.ucar.edu/mailman/listinfo/ncl-talk>
>> Colin M. Zarzycki, ASP postdoctoral fellow
>> <http://mailman.ucar.edu/mailman/listinfo/ncl-talk>
>> Atmospheric Modeling and Predictability
>> National Center for Atmospheric Research
>> <http://mailman.ucar.edu/mailman/listinfo/ncl-talk>
>> http://www.colinzarzycki.com
>>
> _______________________________________________
> ncl-talk mailing list
> List instructions, subscriber options, unsubscribe:
> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>
>
>
> _______________________________________________
> ncl-talk mailing list
> List instructions, subscriber options, unsubscribe:
> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucar.edu/pipermail/ncl-talk/attachments/20141013/0dfb646a/attachment.html
More information about the ncl-talk
mailing list