[MITgcm-devel] new r4 SingleCpuIO efficiency

[Dimitris Menemenlis]

Jean-Michel, I have run the current configuration of CS510 with latest  
code and with c61m on Columbia22.
c61m required 53 minutes for 15 days of integration and latest code  
required 54 minutes.
The 1-minute difference is not significant because I/O is shared with  
other jobs and therefore
the exact throughput of I/O depends on many factors that I cannot  
control.

D.

On May 22, 2009, at 7:51 AM, Jean-Michel Campin wrote:

> Hi Dimitris,
>
> If by chance you have time to check with a short run
> how the new version of SingleCpuIO is doing on cs-510
> (well, given the bug I put & remove only yesterday, I guess
> you did not try yet), I would be curious to know if you
> can see a speed up with large amount of r4 io (compared to
> checkpoint61m). The volume of MPI send/receive for r4 io
> has been reduced by 2 (was r8 before) and this might speed up
> a little the run. And if something goes wrong, let me know.
>
> Thanks,
> Jean-Michel