I mean independent open

[William Gropp]

At 11:35 PM 2/5/2004, Roger Ting wrote:
>Sorry mistake. I mean independent open.

The collective open and collective read/write provide the basis for much of 
the performance advantage that pnetcdf has over non-collective 
approaches.  For individual processes accessing files, there is 
MPI_COMM_SELF, as Rob mentions.  If you want an arbitrary collection of 
processes to access the file, without any a priori coordination (which 
collective open gives you), you open Pandora's box of difficult 
synchronization and consistency issues.  In such cases, it might make more 
sense to use a two stage process, where each process writes a separate file 
(thus eliminating any consistency issues) together with metadata about what 
it wrote, followed by a step that uses collective I/O calls to assemble a 
single netCDF file with maximum efficiency.

Bill



>On Fri, 2004-02-06 at 16:32, Roger Ting wrote:
> > Hi
> >   It seems to be a pity that parallel netcdf doesn't support collective
> > open. It is useful in complex situation when you cannot anticipate which
> > processor and when will a netcdf file is opened.
> >   Any way around this problem beside using a semaphore or condition
> > variable?
> >
> > Roger
> >