On Wed, 2007-21-03 at 09:32 -0700, Bill Loewe wrote:
> 
> Hi Brian,

Hi Bill,

> I handle most of the development/maintenance of IOR, and I'm not sure why the 
> "pass1" output seems to be showing serialized behavior.

That would be mpich1 I guess.  Unlike say pdsh, which threads out for
each of it's remote shells in parallel (to the limit of the fan-out
anyway), I guess mpirun does it more serially.  It seems silly, indeed,
but I have not torn mpich1 apart to discover why so I will refrain from
judging.  :-)

> At a minimum, the 
> source of mpi_time.c

mpi_time.c?

> would need a call to TimeDeviation() in IOR.c.  This 
> function calculates the time offset for each rank against rank 0.  Without 
> the call, there is no adjustment for clock skew -- or rather, the adjustment 
> is always by 0 for all tasks.

Sounds right.

> The latest version of IOR contains Weikuan's patches (massaged to remove the 
> redundancy you noted) if you want to try working with that.  Let me know if 
> you'd like a tarball of this version (the ftp website is not available right 
> now).

Sure, send that along please.

Thanx!

b.



_______________________________________________
Lustre-discuss mailing list
[email protected]
https://mail.clusterfs.com/mailman/listinfo/lustre-discuss

Reply via email to