Exciting.. SOLVED.. There is an open ticket #2043 regarding Nehelem/OpenMPI/Hang problem ( Seems like the problem might be specific to gcc4.4x and OMPI <1.3.2.. It seems like there is a group up us with dual socket nehalems trying to use ompi without much luck (or at least not without headaches).. 

Of note, mca btl_sm_num_fifos 7 seems to work as well..

now off to see if I can get some real code to work... 

Thanks, Mark, Gus, and the rest of the OMPI Users Group!

On Dec 10, 2009, at 7:42 AM, Mark Bolstad wrote:

Just a quick interjection, I also have a dual-quad Nehalem system, HT on, 24GB ram, hand compiled 1.3.4 with options: --enable-mpi-threads --enable-mpi-f77=no --with-openib=no

With v1.3.4 I see roughly the same behavior, hello, ring work, connectivity fails randomly with np >= 8. Turning on -v increased the success, but still hangs. np = 16 fails more often, and the hang is random in which pair of processes are communicating.

However, it seems to be related to the shared memory layer problem. Running with -mca btl ^sm works consistently through np = 128.

Hope this helps.


On Wed, Dec 9, 2009 at 8:03 PM, Gus Correa <> wrote:
Hi Matthew

Save any misinterpretation I may have made of the code:

Hello_c has no real communication, except for a final Barrier
Each process prints "hello world" and that's it.

Ring probes a little more, with processes Send(ing) and
Recv(cieving) messages.
Ring just passes a message sequentially along all process
ranks, then back to rank 0, and repeat the game 10 times.
Rank 0 is in charge of counting turns, decrementing the counter,
and printing that (nobody else prints).
With 4 processes:
0->1->2->3->0->1... 10 times

In connectivity every pair of processes exchange a message.
Therefore it probes all pairwise connections.
In verbose mode you can see that.

These programs shouldn't hang at all, if the system were sane.
Actually, they should even run with a significant level of
oversubscription, say,
-np 128  should work easily for all three programs on a powerful
machine like yours.



1) Stick to the OpenMPI you compiled.


2) You can run connectivity_c in verbose mode:

home/macmanes/apps/openmpi1.4/bin/mpirun -np 8 connectivity_c -v

(Note the trailing "-v".)

It should tell more about who's talking to who.


3) I wonder if there are any BIOS settings that may be required
(and perhaps not in place) to make the Nehalem hyperthreading to
work properly in your computer.

You reach the BIOS settings by typing <DEL> or <F2>
when the computer boots up.
The key varies by
BIOS and computer vendor, but shows quickly on the bootup screen.

You may ask the computer vendor about the recommended BIOS settings.
If you haven't done this before, be careful to change and save only
what really needs to change (if anything really needs to change),
or the result may be worse.
(Overclocking is for gamers, not for genome researchers ... :) )


4) What I read about Nehalem DDR3 memory is that it is optimal
on configurations that are multiples of 3GB per CPU.
Common configs. in dual CPU machines like yours are
6, 12, 24 and 48GB.
The sockets where you install the memory modules also matter.

Your computer has 20GB.
Did you build the computer or upgrade the memory yourself?
Do you know how the memory is installed, in which memory sockets?
What does the vendor have to say about it?

See this:


5) As I said before, typing "f" then "j" on "top" will add
a column (labeled "P") that shows in which core each process is running.
This will let you observe how the Linux scheduler is distributing
the MPI load across the cores.
Hopefully it is load-balanced, and different processes go to different


It is very disconcerting when MPI processes hang.
You are not alone.
The reasons are not always obvious.
At least in your case there is no network involved or to troubleshoot.


I hope it helps,

Gus Correa
Gustavo Correa
Lamont-Doherty Earth Observatory - Columbia University
Palisades, NY, 10964-8000 - USA

Matthew MacManes wrote:
Hi Gus and List,

1st of all Gus, I want to say thanks.. you have been a huge help, and when I get this fixed, I owe you big time!

However, the problems continue...

I formatted the HD, reinstalled OS to make sure that I was working from scratch.  I did your step A, which seemed to go fine:

macmanes@macmanes:~$ which mpicc
macmanes@macmanes:~$ which mpirun

Good stuff there...

I then compiled the example files:

macmanes@macmanes:~/Downloads/openmpi-1.4/examples$ /home/macmanes/apps/openmpi1.4/bin/mpirun -np 8 ring_c
Process 0 sending 10 to 1, tag 201 (8 processes in ring)
Process 0 sent to 1
Process 0 decremented value: 9
Process 0 decremented value: 8
Process 0 decremented value: 7
Process 0 decremented value: 6
Process 0 decremented value: 5
Process 0 decremented value: 4
Process 0 decremented value: 3
Process 0 decremented value: 2
Process 0 decremented value: 1
Process 0 decremented value: 0
Process 0 exiting
Process 1 exiting
Process 2 exiting
Process 3 exiting
Process 4 exiting
Process 5 exiting
Process 6 exiting
Process 7 exiting
macmanes@macmanes:~/Downloads/openmpi-1.4/examples$ /home/macmanes/apps/openmpi1.4/bin/mpirun -np 8 connectivity_c
Connectivity test on 8 processes PASSED.
macmanes@macmanes:~/Downloads/openmpi-1.4/examples$ /home/macmanes/apps/openmpi1.4/bin/mpirun -np 8 connectivity_c

this is maddening because ring_c works.. and connectivity_c worked the 1st time, but not the second... I did it 10 times, and it worked twice.. here is the TOP screenshot:

What is the difference between connectivity_c and ring_c? Under what circumstances should one fail and not the other...

I'm off to the Linux forums to see about the Nehalem kernel issues..


On Wed, Dec 9, 2009 at 13:25, Gus Correa < <>> wrote:

   Hi Matthew

   There is no point in trying to troubleshoot MrBayes and ABySS
   if not even the OpenMPI test programs run properly.
   You must straighten them out first.




   A) While you are at OpenMPI, do yourself a favor,
   and install it from source on a separate directory.
   Who knows if the OpenMPI package distributed with Ubuntu
   works right on Nehalem?
   Better install OpenMPI yourself from source code.
   It is not a big deal, and may save you further trouble.


   1) Install gfortran and g++ if you don't have them using apt-get.
   2) Put the OpenMPI tarball in, say /home/matt/downolads/openmpi
   3) Make another install directory *not in the system directory tree*.
   Something like "mkdir /home/matt/apps/openmpi-X.Y.Z/" (X.Y.Z=version)
   will work
   4) cd /home/matt/downolads/openmpi
   5) ./configure CC=gcc CXX=g++ F77=gfortran FC=gfortran  \
   (Use the prefix flag to install in the directory of item 3.)
   6) make
   7) make install
   8) At the bottom of your /home/matt/.bashrc or .profile file
   put these lines:

   export PATH=/home/matt/apps/openmpi-X.Y.Z/bin:${PATH}
   export MANPATH=/home/matt/apps/openmpi-X.Y.Z/share/man:`man -w`

   (If you use csh/tcsh use instead:
   setenv PATH /home/matt/apps/openmpi-X.Y.Z/bin:${PATH}

   9) Logout and login again to freshen um the environment variables.
   10) Do "which mpicc"  to check that it is pointing to your newly
   installed OpenMPI.
   11) Recompile and rerun the OpenMPI test programs
   with 2, 4, 8, 16, .... processors.
   Use full path names to mpicc and to mpirun,
   if the change of PATH above doesn't work right.


   B) Nehalem is quite new hardware.
   I don't know if the Ubuntu kernel 2.6.31-16 fully supports all
   of Nehalem features, particularly hyperthreading, and NUMA,
   which are used by MPI programs.
   I am not the right person to give you advice about this.
   I googled out but couldn't find a clear information about
   minimal kernel age/requirements to have Nehalem fully supported.
   Some Nehalem owner in the list could come forward and tell.


   C) On the top screenshot you sent me, please try it again
   (after you do item A) but type "f" and "j" to show the processors
   that are running each process.


   D) Also, the screeshot shows 20GB of memory.
   This sounds not as a optimal memory for Nehalem,
   which tend to be 6GB, 12GB, 24GB, 48GB.
   Did you put together the system, or upgraded the memory yourself,
   of did you buy the computer as is?
   However, this should not break MPI anyway.


   E) Answering your question:
   It is true that different flavors of MPI
   used to compile (mpicc) and run (mpiexec) a program would probably
   break right away, regardless of the number of processes.
   However, when it comes to different versions of the
   same MPI flavor (say OpenMPI 1.3.4 and OpenMPI 1.3.3)
   I am not sure it will break.
   I would guess it may run but not in a reliable way.
   Problems may appear as you stress the system with more cores, etc.
   But this is just a guess.


   I hope this helps,

   Gus Correa
   Gustavo Correa
   Lamont-Doherty Earth Observatory - Columbia University
   Palisades, NY, 10964-8000 - USA

   Matthew MacManes wrote:

       Hi Gus,

       Interestingly the results for the connectivity_c test... works
       fine with -np <8. For -np >8 it works some of the time, other
       times it HANGS. I have got to believe that this is a big clue!!
       Also, when it hangs, sometimes I get the message "mpirun was
       unable to cleanly terminate the daemons on the nodes shown
       below" Note that NO nodes are shown below.   Once, I got -np 250
       to pass the connectivity test, but I was not able to replicate
       this reliable, so I'm not sure if it was a fluke, or what.  Here
       is a like to a screenshop of TOP when connectivity_c is hung
       with -np 14.. I see that 2 processes are only at 50% CPU usage..

       The other tests, ring_c, hello_c, as well as the cxx versions of
       these guys with with all values of -np.

       Using -mca mpi-paffinity_alone 1 I get the same behavior.
       I agree that I am should worry about the mismatch between where
       the libraries are installed versus where I am telling my
       programs to look for them. Would this type of mismatch cause
       behavior like what I am seeing, i.e. working with  a small
       number of processors, but failing with larger?  It seems like a
       mismatch would have the same effect regardless of the number of
       processors used. Maybe I am mistaken. Anyway, to address this,
       which mpirun gives me /usr/local/bin/mpirun.. so to configure
       ./configure --with-mpi=/usr/local/bin/mpirun and to run
       /usr/local/bin/mpirun -np X ...  This should
       uname -a gives me: Linux macmanes 2.6.31-16-generic #52-Ubuntu
       SMP Thu Dec 3 22:07:16 UTC 2006 x86_64 GNU/Linux


       On Dec 8, 2009, at 8:50 PM, Gus Correa wrote:

           Hi Matthew

           Please see comments/answers inline below.

           Matthew MacManes wrote:

               Hi Gus, Thanks for your ideas.. I have a few questions,
               and will try to answer yours in hopes of solving this!!

           A simple way to test OpenMPI on your system is to run the
           test programs that come with the OpenMPI source code,
           hello_c.c, connectivity_c.c, and ring_c.c:

           Get the tarball from the OpenMPI site, gzip and untar it,
           and look for it in the "examples" directory.
           Compile it with /your/path/to/openmpi/bin/mpicc hello_c.c
           Run it with /your/path/to/openmpi/bin/mpiexec -np X a.out
           using X = 2, 4, 8, 16, 32, 64, ...

           This will tell if your OpenMPI is functional,
           and if you can run on many Nehalem cores,
           even with oversubscription perhaps.
           It will also set the stage for further investigation of your
           actual programs.

               Should I worry about setting things like --num-cores
               --bind-to-cores?  This, I think, gets at your questions
               about processor affinity.. Am I right? I could not
               exactly figure out the -mca mpi-paffinity_alone stuff...

           I use the simple minded -mca mpi-paffinity_alone 1.
           This is probably the easiest way to assign a process to a core.
           There more complex  ways in OpenMPI, but I haven't tried.
           Indeed, -mca mpi-paffinity_alone 1 does improve performance of
           our programs here.
           There is a chance that without it the 16 virtual cores of
           your Nehalem get confused with more than 3 processes
           (you reported that -np > 3 breaks).

           Did you try adding just -mca mpi-paffinity_alone 1  to
           your mpiexec command line?

               1. Additional load: nope. nothing else, most of the time
               not even firefox.

           Turn off firefox, etc, to make it even better.
           Ideally, use runlevel 3, no X, like a computer cluster node,
           but this may not be required.

               2. RAM: no problems apparent when monitoring through
               TOP. Interesting, I did wonder about oversubscription,
               so I tried the option --nooversubscription, but this
               gave me an error mssage.

           Oversubscription from your program would only happen if
           you asked for more processes than available cores, i.e.,
           -np > 8 (or "virtual" cores, in case of Nehalem hyperthreading,
           -np > 16).
           Since you have -np=4 there is no oversubscription,
           unless you have other external load (e.g. Matlab, etc),
           but you said you don't.

           Yet another possibility would be if your program is threaded
           (e.g. using OpenMP along with MPI), but considering what you
           said about OpenMP I would guess the programs don't use it.
           For instance, you launch the program with 4 MPI processes,
           and each process decides to start, say, 8 OpenMP threads.
           You end up with 32 threads and 8 (real) cores (or 16
           ones on Nehalem).

           What else does top say?
           Any hog processes (memory- or CPU-wise)
           besides your program processes?

               3. I have not tried other MPI flavors.. Ive been
               speaking to the authors of the programs, and they are
               both using openMPI.  

           I was not trying to convince you to use another MPI.
           I use MPICH2 also, but OpenMPI reigns here.
           The idea or trying it with MPICH2 was just to check whether
           is causing the problem, but I don't think it is.

               4. I don't think that this is a problem, as I'm
               specifying --with-mpi=/usr/bin/...  when I compile the
               programs. Is there any other way to be sure that this is
               not a problem?

           Hmmm ....
           I don't know about your Ubuntu (we have CentOS and Fedora on
           However, most Linux distributions come with their MPI flavors,
           and so do compilers, etc.
           Often times they install these goodies in unexpected places,
           and this has caused a lot of frustration.
           There are tons of postings on this list that eventually
           boiled down to mismatched versions of MPI in unexpected places.

           The easy way is to use full path names to compile and to run.
           Something like this:
           /my/openmpi/bin/mpicc on your program configuration script),

           and something like this
           /my/openmpi/bin/mpiexec -np  ... bla, bla ...
           when you submit the job.

           You can check your version with "which mpicc", "which mpiexec",
           and (perhaps using full path names) with
           "ompi_info", "mpicc --showme", "mpiexec --help".

               5. I had not been, and you could see some shuffling when
               monitoring the load on specific processors. I have tried
               to use --bind-to-cores to deal with this. I don't
               understand how to use the -mca options you asked about.
               6. I am using Ubuntu 9.10. gcc 4.4.1 and g++  4.4.1

           I am afraid I won't be of help, because I don't have Nehalem.
           However, I read about Nehalem requiring quite recent kernels
           to get all of its features working right.

           What is the output of "uname -a"?
           This will tell the kernel version, etc.
           Other list subscribers may give you a suggestion if you post the

               MyBayes is a for bayesian phylogenetics:
               ABySS: is a program for assembly of DNA sequence data:

           Thanks for the links!
           I had found the MrBayes link.
           I eventually found what your ABySS was about, but no links.
           Amazing that it is about DNA/gene sequencing.
           Our abyss here is the deep ocean ... :)
           Abysmal difference!

                   Do the programs mix MPI (message passing) with
                   OpenMP (threads)?

               Im honestly not sure what this means..

           Some programs mix the two.
           OpenMP only works in a shared memory environment (e.g. a single
           computer like yours), whereas MPI can use both shared memory
           and work across a network (e.g. in a cluster).
           There are other differences too.

           Unlikely that you have this hybrid type of parallel program,
           otherwise there would be some reference to OpenMP
           on the very program configuration files, program
           documentation, etc.
           Also, in general the configuration scripts of these hybrid
           programs can turn on MPI only, or OpenMP only, or both,
           depending on how you configure.

           Even to compile with OpenMP you would need a proper compiler
           flag, but that one might be hidden in a Makefile too, making
           a bit hard to find. "grep -n mp Makefile" may give a clue.
           Anything on the documentation that mentions threads or OpenMP?

           FYI, here is OpenMP:

               Thanks for all your help!

            > Matt

           Well, so far it didn't really help. :(

           But let's hope to find a clue,
           maybe with a little help of
           our list subscriber friends.

           Gus Correa
           Gustavo Correa
           Lamont-Doherty Earth Observatory - Columbia University
           Palisades, NY, 10964-8000 - USA

                   Hi Matthew

                   More guesses/questions than anything else:

                   1) Is there any additional load on this machine?
                   We had problems like that (on different machines) when
                   users start listening to streaming video, doing
                   Matlab calculations,
                   etc, while the MPI programs are running.
                   This tends to oversubscribe the cores, and may lead
                   to crashes.

                   2) RAM:
                   Can you monitor the RAM usage through "top"?
                   (I presume you are on Linux.)
                   It may show unexpected memory leaks, if they exist.

                   On "top", type "1" (one) see all cores, type "f"
                   then "j"
                   to see the core number associated to each process.

                   3) Do the programs work right with other MPI flavors
                   (e.g. MPICH2)?
                   If not, then it is not OpenMPI's fault.

                   4) Any possibility that the MPI versions/flavors of
                   mpicc and
                   mpirun that you are using to compile and launch the
                   program are not the

                   5) Are you setting processor affinity on mpiexec?

                   mpiexec -mca mpi_paffinity_alone 1 -np ... bla, bla ...

                   Context switching across the cores may also cause
                   trouble, I suppose.

                   6) Which Linux are you using (uname -a)?

                   On other mailing lists I read reports that only
                   quite recent kernels
                   support all the Intel Nehalem processor features well.
                   I don't have Nehalem, I can't help here,
                   but the information may be useful
                   for other list subscribers to help you.


                   As for the programs, some programs require specific
                   (and even specific compilation) when the number of
                   MPI processes
                   It may help if you tell us a link to the program sites.

                   Baysian statistics is not totally out of our business,
                   but phylogenetic genetic trees is not really my league,
                   hence forgive me any bad guesses, please,
                   but would it need specific compilation or a different
                   set of input parameters to run correctly on a different
                   number of processors?
                   Do the programs mix MPI (message passing) with
                   OpenMP (threads)?

                   I found this MrBayes, which seems to do the above:


                   As for the ABySS, what is it, where can it be found?
                   Doesn't look like a deep ocean circulation model, as
                   the name suggest.

                   My $0.02
                   Gus Correa

               users mailing list


           users mailing list


       Matthew MacManes
       PhD Candidate
       University of California- Berkeley
       Museum of Vertebrate Zoology
       Phone: 510-495-5833
       Lab Website:
       Personal Website:


       users mailing list <>

   users mailing list <>


users mailing list

users mailing list

users mailing list

Matthew MacManes
PhD Candidate
University of California- Berkeley
Museum of Vertebrate Zoology
Phone: 510-495-5833
Lab Website:
Personal Website: