On Fri, 20 Apr 2007, João Luis Silva wrote:
I´m new to the list, please excuse if this is a faq:
Has anyone experience with meep-mpi? What speedup can be expected?
(I get no speed up at all on my dual core laptop.)
My institute plans to buy a small cluster machine, are there any
suggestions for configuration (RAM, CPU, interconnect)?
Thanks,
I have used meep-mpi a few times, and mostly because the problem would not
fit the memory of a single computer. I tried to understand how it would
scale in terms of CPU without much success. One thing you should do is
figure out how much RAM the problems you want to solve require, and ensure
the cluster total memory is at least this much. The cluster I used consists
of some 21 dual opterons with 4GB RAM per machine, and gigabit ethernet.
Maybe someone with more experience than me will be able to clarify this
further.
Basically, the key to getting a good speedup is to make sure that:
1) the problem is not too small (otherwise the communication costs will
outweigh the computational benefits)
2) there isn't too much I/O (I/O tends not to parallelize too well unless
you spend a lot of effort on the hardware end)
For example, I took the ring.ctl example file that is included with Meep,
and commented out the output-epsilon and output-efield-z lines to remove
the HDF5 output. I also increased the resolution to 20 to make the
problem a little bigger (otherwise it runs in < 2 seconds). Then, on an
3GHz Intel Core Duo, it takes 10 seconds on 1 CPU, and 6.8 seconds on 2
CPUs. This is still a rather small problem; as the problem sizes
increases, the speedup improves.
Steven
_______________________________________________
meep-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/meep-discuss
