nfs opinion

[Bohn Christopher A Capt AFRL/IFSD]

The advice I gave to my cohorts was to read & write to local disk if they
were going to have much disk access in the middle of execution (vice
beginning & end only) and then collect the results at the end.  Obviously
not an answer for diskless nodes...

> -----Original Message-----
> From: Peter de Groot [mailto:pdgtech@kalnet.com.au]
> Sent: Wednesday, June 16, 1999 1:04 PM
> To: beowulf@beowulf.gsfc.nasa.gov
> Cc: Martin de Groot; Tony de Groot
> Subject: nfs opinion
> 
> 
> I am beginning the come to the conclusion that there
> needs to be a fundimental re-think with nfs, with regards
> to file locking, caching etc.
> 
> With a number of nodes banging away against a file system,
> concurrency is becoming a major hassle.
> 
> Our software vendor creates and then deletes  so called
> "lock" file in the target directory, when opening data files 
> for write. 
>  But with caching, this may be a moot point..... to set the
> mount as nocache would kill nfs performance.  Should we create
> and export nocache file system for lock files ??? yuk.
> 
> I know there is something around called lockd.  But I think
> that that you have to write against it in the software.
> 
> I reckon this is going to be a problem as Beowulfs try
> to become more "homogeneous"
> 
> Peter
> 
> 
> 
> 
> 
> > 
> > It is required for correctness.  If noac is not set, NFS 
> caches data about
> > files without maintaining coherence.  It is a tradeoff (in 
> NFS) between
> > speed and correctness.  For many applications where any 
> file is being
> > access by only a single process on a single node, there aren't any
> > problems.  Once you have multiple processes on different 
> nodes accessing
> > the file, whether they are running MPI or Emacs, you can 
> run into problems
> > if noac is not set.  More specifically for MPI, the MPI I/O 
> rules require
> > that changes to a file being accessed from within a 
> parallel program meet
> > some relatively loose consistency rules.  NFS, without 
> noac, does not meet
> > these rules.
> > 
> > Bill
> 
> -- 
> 
> 
> ______________________________________________________________
> 
>               PDG Technical Services Pty Ltd        
>  Peter de Groot                             P.O. Box 10349   
>   08) 90916817                              Kalgoorlie 6430  
>                                             Western Australia
> ______________________________________________________________
>