[mpich-discuss] MPI_File_set_view Error: ADIOI_Count_contiguous_blocks

[Rajeev Thakur]

> Uhmm, no idea, regular NFS PC hard drives? (All are blade clusters.)

You won't get much performance if you use NFS-mounted home directories.

Rajeev 

> -----Original Message-----
> From: owner-mpich-discuss at mcs.anl.gov 
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Ingo Bojak
> Sent: Friday, May 30, 2008 11:23 AM
> To: mpich-discuss at mcs.anl.gov
> Subject: Re: [mpich-discuss] MPI_File_set_view Error: 
> ADIOI_Count_contiguous_blocks
> 
> Robert Latham wrote:
> > - How many processes are writing?
> All processes in the ring, i.e., typically no less than 8 and no more 
> than about 40.
> > - About how much data is each process writing?
> >   
> Two scenarios:
> 1) in single state variable tracking about (20,000 to 70,000) doubles 
> repeatedly, data evenly spread over nodes.
> 2) in checkpoint dumping about (20,000 to 70,000) x 70 doubles, once, 
> also evenly spread
> > - Is the data noncontiguous in memory? file? both?
> >   
> The data is noncontiguous in the file (complicated write pattern set 
> with fileview for every process). The dara is noncontiguous 
> in memory, 
> actually transpose to what would be needed for write out, so 
> I currently 
> transpose it into a contiguous write-out buffer for the 
> writing with the 
> file view.
> > - What's the interconnect for this machine?
> >   
> Good question, since it will run on several different 
> clusters, and I'm 
> not terribly  familiar with the details of all of them. I 
> think Gigabit 
> ethernet on one, and 20 Gb/s Infiniband on the other.
> > - How fast is your storage?
> >   
> Uhmm, no idea, regular NFS PC hard drives? (All are blade clusters.)
> 
> Cheers,
> Ingo
> 
> 
>