cl-mpi and debian 8

[quicklisp]

I'm trying to build on debian 8. It installs mpi.h in /usr/lib/openmpi/openmpi/include/mpi.h. What should I do to help cl-openmpi find it?

[marcoheisig]

Greetings!

The short answer:
$ sudo update-alternatives --set mpi /usr/lib/openmpi/include
$ sudo update-alternatives --set mpirun /usr/bin/mpirun.openmpi
The former command places a C compiler wrapper script called mpicc in the $PATH which cl-mpi uses instead of $CC to grovel all MPI constants and load all MPI libraries. The latter command makes the mpirun program accessible which is used to launch cl-mpi images in parallel.

The long answer:
There is no sane way of automatically configuring and detecting all include paths, libraries and flags necessary to use the active MPI installation of a computer - I tried and failed. What the C guys do instead is to use a wrapper script called mpicc that calls the host compiler with all necessary parameters, e.g. where mpi.h is located. Each MPI library like OpenMPI and MPICH provides their own suitable mpicc command. I use mpicc for CFFI grovelling and to build a small shared library called cl-mpi-stub.so which I then load from CFFI to pull in all library dependencies. This approach seems to work well on all systems I tested so far.

I hope this solves the problem. My development platform is also Debian 8, so I am optimistic. If not I would like to see the output of
$ mpicc -show
and the error message of (asdf:load-system 'cl-mpi).

Best regards,
Marco

[quicklisp]

Thanks.

I get this:

; gcc -m64 -I/home/quicklisp/quicklisp-controller/dist/build-cache/cffi/e7921b040f0a965a2604c95da10e039387059c22/cffi_0.16.1/ -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/a8c717e7b6f8da88a38d108da10fb6c2ab710231/cl-mpi-20160110-git/mpi/grovel /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/a8c717e7b6f8da88a38d108da10fb6c2ab710231/cl-mpi-20160110-git/mpi/grovel.c
A nested error within --disable-debugger error handling prevents displaying the original error. Attempting to print a backtrace.

When I run that from the command line, I get this:

$ gcc -m64 -I/home/quicklisp/quicklisp-controller/dist/build-cache/cffi/e7921b040f0a965a2604c95da10e039387059c22/cffi_0.16.1/ -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/a8c717e7b6f8da88a38d108da10fb6c2ab710231/cl-mpi-20160110-git/mpi/grovel /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/a8c717e7b6f8da88a38d108da10fb6c2ab710231/cl-mpi-20160110-git/mpi/grovel.c
/home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/a8c717e7b6f8da88a38d108da10fb6c2ab710231/cl-mpi-20160110-git/mpi/grovel.c:6:17: fatal error: mpi.h: No such file or directory
 #include <mpi.h>
         ^
compilation terminated.

And the mpicc output:

$ mpicc -show
gcc -I/usr/lib/openmpi/include -I/usr/lib/openmpi/include/openmpi -pthread -L/usr//lib -L/usr/lib/openmpi/lib -lmpi -ldl -lhwloc

[marcoheisig]

Thank you for the helpful output!

The output on the command line should begin with $ mpicc -m64 ... and I think I know why it does not. I used the following code via ASDFs :defsystem-depends-on to force CFFI to use mpicc instead of gcc for grovelling.

(defmethod perform :around ((op cffi-grovel::process-op)
                            (c cffi-grovel:grovel-file))
  (let ((cffi-grovel::*cc* "mpicc"))
    (call-next-method)))

It appears another :around method has replaced this one in your case, restoring the original behavior. This is of course my fault. I already uploaded a fix that subclasses cffi-grovel:grovel-file for MPI related operations. I hope your problem is now gone.

[quicklisp]

Got the new commit, seeing the same issue. Here's more output:

; mpicc -shared -fPIC -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.so /home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.c
; gcc -m64 -I/home/quicklisp/quicklisp-controller/dist/build-cache/cffi/e7921b040f0a965a2604c95da10e039387059c22/cffi_0.16.1/ -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel.c
A nested error within --disable-debugger error handling prevents displaying the original error. Attempting to print a backtrace.
cl-mpi-extensions cl-mpi-extensions
; mpicc -shared -fPIC -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.so /home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.c
; gcc -m64 -I/home/quicklisp/quicklisp-controller/dist/build-cache/cffi/e7921b040f0a965a2604c95da10e039387059c22/cffi_0.16.1/ -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel.c
A nested error within --disable-debugger error handling prevents displaying the original error. Attempting to print a backtrace.
cl-mpi-testsuite cl-mpi-testsuite
; mpicc -shared -fPIC -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.so /home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.c
; gcc -m64 -I/home/quicklisp/quicklisp-controller/dist/build-cache/cffi/e7921b040f0a965a2604c95da10e039387059c22/cffi_0.16.1/ -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel.c
A nested error within --disable-debugger error handling prevents displaying the original error. Attempting to print a backtrace.
cl-mpi cl-mpi
; mpicc -shared -fPIC -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.so /home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/cl-mpi-stub.c
; gcc -m64 -I/home/quicklisp/quicklisp-controller/dist/build-cache/cffi/e7921b040f0a965a2604c95da10e039387059c22/cffi_0.16.1/ -o /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel /home/quicklisp/.cache/common-lisp/sbcl-1.3.1.188-f7d76e5-linux-x64/home/quicklisp/quicklisp-controller/dist/build-cache/cl-mpi/9a2e10da9fbf0fd92742014efa5ce6e5c170e9c4/cl-mpi-20160111-git/mpi/grovel.c
A nested error within --disable-debugger error handling prevents displaying the original error. Attempting to print a backtrace.

[marcoheisig]

If this does not fix it, i have just discovered that CFFI determines the C compiler with

(or (getenv "CC") *cc*)

so unsetting the environment variable $CC or setting it to mpicc might help.

[quicklisp]

Progress! Unsetting CC worked for cl-mpi. Unfortunately, I think unsetting it breaks some other projects. I'll run through building the universe and see what happens with no CC.

[marcoheisig]

If it helps I can detect if $CC is set and if so, bind it to mpicc for the dynamic extent of the CFFI grovelling and restore it afterwards. Then other projects can use $CC as they want. I only fear changing environment variables (even for such a short duration) could be a horrible source of nondeterminism for automated build systems. What do you think?

[quicklisp]

I unset CC in my build process. cl-mpi started working, and nothing broke. I'm going to leave it unset and consider things fixed. Thanks for your help!