Hi,
On środa, 6 czerwca 2007, sanjib gupta wrote:
> I am attaching 2 log files - the initial run on 128 processors, then
> immediately killing the job and restarting from the first checkpoint
> file "hc-rt-hdf5_chk_0000"
> notice about 4 timesteps per second initially, then ~30 timesteps/sec
> after restart.
You're using only base level of refinement (lrefine_max=1) which has a non
obvious side effect: when you start from scratch, only master procesor gets
the work and the other are just waiting. After a restart the blocks are
distributed and then the run becomes truly parallel. I'd suggest you to
decrease nblock[xy] and use lrefine_min=2 and lrefine_max=2 instead, then
after refining to second level the blocks will be distributed.
If you don't require extremely flexible AMR then you may also increase nxb
and nyb from 8 to 16 or 32 (at compile time) which will reduce overhead due
to communication.
Your setup is also relatively small - it's just 32x128 cells, so probably it
makes little sense to use more than 4 or 8 CPU for it.
--
Cheers,
Artur
Received on Wed Jun 6 00:12:00 2007
This archive was generated by hypermail 2.1.8 : Wed Jun 06 2007 - 00:12:01 CDT