On Wed, 2008-02-13 at 09:49 -0700, Ed Hartnett wrote:
> Thomas Green <thomas.green@xxxxxxxxxxxxxxxx> writes:
>
> > Fellow NetCDF users,
> >
> > I am currently evaluating NetCDF on a number of platforms but I am
> > having difficulty with using NetCDF4 on a NEC SX-6. I follow the
> > instructions and get everything compiled without any problems but I have
> > problems at runtime. (I have tried both parallel and serial versions).
> >
> > I have traced the problem to the new HDF5 library which fails on its own
> > make tests (and I can also use the NetCDF4 library using NetCDF3 classic
> > mode). I have emailed the HDFgroup support but since NEC is not
> > supported by them there is not much they can do.
> >
> > I am really after anyone out in NetCDF world who is on the bleeding edge
> > with NetCDF4 and is on an NEC. The new NetCDF4 capabilities are really
> > useful and I have them running on Linux with no problems.
> >
> > Any pointers welcome. I can provide the support request I emailed
> > help@xxxxxxxxxxxx if people need more details about the error.
> >
> > Many thanks,
> >
> > Tom
>
> Howdy Tom!
>
> As you note, if the HDF5 library won't work, there's not much to be
> done on the netCDF end.
>
> If I were you, I might try and get HDF5 to work on NEC. Frequently,
> porting to a new machine is not that big of a deal.
>
> However, the HDF5 porter is no doubt firing up his email to tell me
> that it can also be very difficult. (Hi Kent!). Of course he is
> correct.
>
> But frequently this task is easiest for a platform specialist, so if
> you know the NEC, you might be the best person to try it.
>
> Thanks!
>
> Ed
Ed,
Many thanks for the vote of confidence on my porting abilities. I might
try and dig a bit more into the source code. My first idea was the io
performance affecting the HDF5 library since if I change the option in
the HDF tests to stress the IO less it seems to get a bit further
(eventually same error though) but it could also be a coincidence. We
currently have to use a large chunksize with NetCDF3 files on the NEC to
realise its full io potential (approx 32mb). I was trying to find out
how HDF5 behaves with io which requires large data transfers so its uses
the faster pathway but I got as far as HDF property lists but didnt have
time to dig much further. I will post here if I get any further.
Thanks again,
Tom
---
Thomas Green Unified Model System Analyst
Met Office FitzRoy Road Exeter EX1 3PB United Kingdom
http://www.metoffice.gov.uk
