Hi Jeff,
Let me try to help
Matthias,
This is a summary of my problems on Jacquard (http://www.nersc.gov/nusers/resources/jacquard/) at NERSC.
Attempt One
- hdf5/1.6.4
HDF5 support is not really implemented yet in the current version, hence it might help to configure with --without-hdf5
configure: error: unknown mode pathscale
The first "out of the box" fix would be support for Pathscale compilers
We can easily add that if you can tell me which compiler options should be default for pathscale? The same as for gnu?
Running configure assuming Pathscale is GNU equivalent jacin01:~/alps-1.3b2> ./configure --prefix=~/ALPS --with-compiler=gnu
It now picks up MPI but no HDF5. I would have though the presence of ACML and SCALAPACK would give me LAPACK somehow but perhaps not.
It just does not find it automatically yet. YOur next step is just what is needed, with some extra thibgs
I added --with-atlas-dir=/usr/common/acts/ATLAS/lib/Linux_Opteron where libatlas.a etc. is. ALPS configure cannot find dgemm because it is called something else:
ar -t /usr/common/acts/ATLAS/lib/Linux_Opteron/libatlas.a |
grep dgemm ATL_dgemm.o ATL_dgemmNN.o ATL_dgemmNT.o ATL_dgemmTN.o ATL_dgemmTT.o ATL_dgemmRB.o
Yes, I know this problem. Usually the ATLAS has a blas library that goes with it which forwards the calls. You need to specify in addition to the atlas dir also the libraries to link against, which will be more than just -latlas. The best way is to try to compil and link a program calling dgemm_. Once you find the needed libraries we can add them to the configure script so that they are tried automatically. Please let me know what is needed.
I attached the config.log from the setup I used to compile with, having given up on LAPACK. The last 50 lines of my make.log are attached. It bombed on MPI and the error requires me to recompile: relocation R_X86_64_32S can not be used when making a shared object; recompile with -fPIC ///usr/common/nsg/mvapich/pathscale/mvapich-0.9.5-mlx1.0.3/lib/ libpmpich++.a: could not read symbols: Bad value
It will really suck if I have to recompile MPI on my own.
The easier option might just to forget about shared libraries and link statically. To do this add the --disable-shared option to the configure call.
Now I can start all over with GNU compilers instead of Pathscale but there are very good reasons to stick with Pathscale on this machine. Please let me know what problems I can correct or what other approaches I should try. It is also helpful to tell me if NERSC should reconfigure any of their libraries.
If it appears the problem is that I am a moron, you're probably right; don't hesitate to say so.
No, not at all. You just have a non-standard setup and we need to figure out what works on your machine. We can then add out-of-the-box support to the next release.
Matthias
comp-phys-alps-users@lists.phys.ethz.ch