possible bug in nfmpi_enddef?

[Rob Latham]

On Fri, May 21, 2010 at 01:23:02PM -0600, Jim Edwards wrote:
> I am trying to do
> 
> rocde = nfmpi_redef(file)
> rcode = nfmpi_put_att(cpl_io_file,pio_global,"file_version",version)
> rcode = nfmpi_enddef(cpl_io_file)

> But this results in
> 
>  Memory allocation (malloc) failure

placing a single attribute on the dataset *might* result in growing
the header.  You'd know it if you ever hit that: performance would go
straight down the toilet.   If the header grows enough that you have
to shuffle all the other variables down, pnetcdf will do so at 4k (!)
at a time.  

So, that's one place that allocates memory in the redef case, but
there really aren't 4k bytes of memory to spare?  Seems unlikely.

Do you maybe have a core file or a backtrace of where this happens?  

Is this on bluefire or bgp or something else?

==rob

-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA