Re: NCL performance

Thanks Mary. It is not critical for us now. If you can resolve the
issue in the next release it would be great. Thanks.

-Fan

Mary Haley wrote:
> Fan,
>
> I was able to verify that the timiing difference happened between versions 5.1.1 and 5.2.0.
> You're right: the issue is with ut_inv_calendar and not the loop.
>
> It's hard to narrow down the problem at the moment, because in 5.2.0, we started using
> Udunits-2, and had to change the code to reflect the new interface. It's possible that Udunits-2
> is the issue, and not NCL, but I'm not certain.
>
> I will submit a ticket on this, but it's going to be a low priority right now.
>
> --Mary
>
> On Nov 30, 2010, at 9:01 AM, Fan Fang wrote:
>
>
>> Hi Mary,
>>
>> I was able to reproduce the problem with the following simple script.
>> It largely resembles what was the most suspicious part of the code that
>> slowed down with 5.2.1. It does a calendar conversion and an assignment
>> in a loop. It doesn't look like that the do-loop itself slowed down.
>>
>> Typically it is a 4-second run for 5.1.1, but >15 seconds for 5.2.1.
>>
>> Thanks.
>>
>> -Fan
>>
>> load "$NCARG_ROOT/lib/ncarg/nclscripts/csm/contributed.ncl"
>> begin
>> x=new(1000,double)
>> y=new(1000,double)
>> wcStrtLooping = systemfunc("date")
>> yyyy = 2002
>> mm = 11
>> dd = 18
>> hh = 5
>> nn = 0
>> ss = 0
>> do i=0,999
>> x@units = "days since 1900-01-01 00:00:00"
>> x(i) = ut_inv_calendar(yyyy,mm,dd,hh,nn,ss,x@units,0)
>> if(i.gt.100) then
>> y(i) = 1.0
>> else
>> y(i) = 0.0
>> end if
>> end do
>> wallClockElapseTime(wcStrtLooping, "Complete loop", 0)
>> end
>>
>>
>> Mary Haley wrote:
>>
>>> Hi Fan,
>>>
>>> We need more information to see if there could be an issue.
>>>
>>> Please insert calls to "wallClockElapseTime in various sections in your script, and then rerun the script
>>> to get more timing results. There are some examples here:
>>>
>>> http://www.ncl.ucar.edu/Document/Functions/Contributed/wallClockElapseTime.shtml
>>>
>>> This will help us narrow down where the problem is. Some key areas, as Dennis pointed out,
>>> could have something to do with file I/O. It would be good to put these calls around
>>> functions that are dealing with large arrays, and any calls to "addfile". Also, any
>>> calls to a graphics routine might be a problem area.
>>>
>>> Thanks,
>>>
>>> --Mary
>>>
>>>
>>> On Nov 29, 2010, at 11:55 AM, Fan Fang wrote:
>>>
>>>
>>>
>>>>> uname -a
>>>>>
>>>>>
>>>> Linux discette 2.6.18-194.11.3.el5PAE #1 SMP Mon Aug 30 17:02:48 EDT
>>>> 2010 i686 i686 i386 GNU/Linux
>>>>
>>>> What I did was to install the two versions (below) at the same location,
>>>> and doing the same command-line (same input and output paths) using the
>>>> same script. I added "time" in front the command-line. Results below.
>>>> Let me know if you need more info. Thanks.
>>>>
>>>> -Fan
>>>>
>>>> ncl_ncarg-5.2.1.Linux_RedHat_i686_nodap_gcc412.tar.gz:
>>>> 172.341u 1.212s 2:54.65 99.3% 0+0k 0+0io 0pf+0w
>>>>
>>>> ncl_ncarg-5.1.1.Linux_i686_gcc412.tar.gz:
>>>> 50.035u 0.579s 0:50.87 99.4% 0+0k 0+0io 0pf+0w
>>>>
>>>> Dennis Shea wrote:
>>>>
>>>>
>>>>> There should be no performance difference.
>>>>> All 5.2.1 updates are listed at:
>>>>> http://www.ncl.ucar.edu/current_release.shtml
>>>>>
>>>>> [1] If your files are on a NFS mounted file system, perhaps the
>>>>> network was having issues.
>>>>>
>>>>> [2] Whenever, you post a comment or question like this one,
>>>>> please include the output from
>>>>>
>>>>> %> uname -a
>>>>>
>>>>> THX
>>>>>
>>>>> On 11/29/10 9:12 AM, Fan Fang wrote:
>>>>>
>>>>>
>>>>>
>>>>>> With NCL 5.2.1, we noticed a factor of ~3 reduction in performance with
>>>>>> a x-y plot script. Our previous (and better performed) NCL version was
>>>>>> 5.1.1. Is this expected?
>>>>>>
>>>>>> The script is plotting time series. There are quite a bit array
>>>>>> manipulations in script. We haven't yet compared other scripts.
>>>>>>
>>>>>> -Fan
>>>>>> _______________________________________________
>>>>>> 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
>>>>
>>>>
>>>
>> _______________________________________________
>> 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
Received on Thu Dec 2 05:23:59 2010