example timings

[michael bane]

On Wed, 2007-07-18 at 12:04 -0500, Wei-keng Liao wrote:
> Michael,
> 
> I noticed that your codes write the output files to your home directory 
> which is NFS mounted. Usually, NFS performance for parallel I/O is very 
> slow.


Wei-keng, I'm confused as to your assertion that my home filestore is
NFS mounted. I can't see how one can tell from the code or graph! I'll
check with the sysAdmin tomorrow but it seems to me that both
directories (for executable and output files) are Lustre:

~$ df -h ~/Fortran/MPI/parallel_netCDF/ ~/tmp/test_par_netcdf/
Filesystem            Size  Used Avail Use% Mounted on
1 at elan0:/mdt_horace1.nec.1/clientelan
                      1.5T  283G  1.2T  20% /home/horace
1 at elan0:/mdt_horace1.nec.1/clientelan
                      1.5T  283G  1.2T  20% /home/horace

~$ mount | grep -e 1 at elan0:/mdt_horace1.nec.1/clientelan
1 at elan0:/mdt_horace1.nec.1/clientelan on /home/horace type lustre
(rw,user_xattr,acl)

~$ 

> Can you try write to the Lustre file system? I can see Lustre is available 
> on your machine. You may also want to configure a Lustre directory with 
> multiple stripe counts to increase the I/O performance.

I've no idea how Lustre has been set up and have forwarded your comments
to our sysAdmin for comment.

Thanks, Michael