[Ipopt] Faster optimization

[Gerry Chu]

Hi Everybody

I'm trying to use ipopt to do real-time robot control. Currently my 
control loop with ipopt runs at 15Hz, and I'd like to make that faster 
if possible. There seems to be a lot of ways to do this, and I need some 
help to direct my efforts into the most fruitful way. I'm using ipopt 
from C.

Solver: I'm using MUMPS/METIS. When I run Ipopt, it suggests using 
another solver for speed. I tried paradiso, but that when I try to

../configure \
--with-pardiso="/usr/local/lib/libpardiso_GNU43_IA32_P.so"    \
--with-blas="/usr/lib/sse2/atlas/libblas.so.3gf"     \
--with-lapack="/usr/lib/sse2/atlas/liblapack.so.3gf"

I get "error: Pardiso library /usr/local/lib/libpardiso_GNU43_IA32_P.so 
does not seem to work", even though the file exists. I also tried using 
the intel mkl for BLAS, but I can't find an example of how to configure it.

Multiprocessing: I have a quad-core Intel core processor, and ipopt uses 
one processor to 100% and none of the others. Can it use other 
processors too?

Warm start: I think my problem would really benefit from warm starts, 
however, I can't find anywhere in the documentation of how to do this. I 
read: http://www.coin-or.org/Ipopt/documentation/node128.html but it is 
poorly explained. Is there an example?

Other ipopt options: there seems to be a lot of ways to tweak the 
solver, however I don't really have a math background and I have no idea 
where to start.

Any suggestions are appreciated.

Thanks

Gerry Chu