[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: 20030627: NEC performance issues (cont.)

>To: address@hidden
>cc: "Ingleby, Bruce" <address@hidden>
>From: Matt Huddleston <address@hidden>
>Subject: Re: NEC performance issues
>Organization: Met Office
>Keywords: 200306271029.h5RATqLd002822 netcdf-3.5.1-beta10  on a NEC SX6

Hi Matt,

> Well, I've solved my performance problems.  Here are two pieces of code
> that do the same thing :
> 
> (A)
> DO Iob=1,NumOb
>   iret = nf90_put_var(ncid,  potm_id, RealNobNLevVar(IOb,:), &
>                                                     start=(/1,IOb/))
> ENDDO
> 
> (B)
> iret = nf90_put_var(ncid, potm_id, RealNobNLevVar,    &
>           count=(/NumObLevUsed,NumOb/), map=(/ NumOb,1 /) )
> 
> 
> And we used code (B) on our Cray T3e as it's more efficient.  However,
> on the NEC SX6, code (B) runs 70 times more slowly than (A)!!!  What are
> those C routines doing to slow the mapping down so much ?!

I'm forwarding the note below, which reports some success dealing with
NEC SX6 performance problems using a new version of netCDF available
from

  ftp://ftp.unidata.ucar.edu/pub/netcdf/exp/netcdf-3.5.1-beta12.tar.Z

If it's convenient, could you build the new version and let us know if
its performance seems adequate for your application?  Thanks!

--Russ

_____________________________________________________________________

Russ Rew                                         UCAR Unidata Program
address@hidden                      http://my.unidata.ucar.edu

------- Forwarded Message

Date:    Tue, 05 Aug 2003 10:46:32 +0200
From:    Mathis Rosenhauer <address@hidden>
To:      Steve Emmerson <address@hidden>
cc:      address@hidden,
         Thomas Schoenemeyer <address@hidden>,
         Gottfried Necker <address@hidden>,
         Russ Rew <address@hidden>, DKRZ Beratung <address@hidden>
Subject: Re: 20030804: Netcdf performance problem on NEC SX6

Hi Steve,

performance is good again on our shared filesystems. I could only
try it with some simple test programs though. Feedback from our users
will take a bit longer but I don't expect complaints.

Thank you very much for your Help

Mathis


On Mon, 4 Aug 2003, Steve Emmerson wrote:

> Mathis,
>
> We've created a new version of the netCDF package that fixes the 1999
> nc_sync() bug but shouldn't have the performance problem that you
> encountered.
>
> Would you be willing to test it?  The new version is available via
> anonymous FTP at
>
>     ftp://ftp.unidata.ucar.edu/pub/netcdf/exp/netcdf-3.5.1-beta12.tar.Z
>
> Please let us know what you decide or discover.
>
> Regards,
> Steve Emmerson
>

-- 
Mathis Rosenhauer
Wissenschaftliches Rechnen
Deutsches Klimarechenzentrum                       http://www.dkrz.de

------- End of Forwarded Message