On May 10, 2006, at 10:46 PM, Gurhan Ozen wrote:

My ultimate goal is to get Open MPI working with openIB stack. First, I had installed lam-mpi , I know it doesn't have support for openIB but it's still relevant to some of my questions I will ask.. Here is the set up I have:

Yes, keep in mind throughout that while Open MPI does support MVAPI, LAM/MPI will fall back to using IP over IB for communication.

I have two machines, pe830-01 and pe830-02 .. Both have ethernet interface and
 HCA interface. The IP addresses follow:
                         eth0                 ib0
 pe830-01     10.12.4.32      192.168.1.32
 pe830-02     10.12.4.34      192.168.1.34

So this has worked even though it lamhosts file is configured to use ib0 interfaces. I further verified with tcpdump command that none of this went
   to eth0 ..

Anyhow, if i change the lamhosts file to use the eth0 IPs, things work just as the same with no issues . And in that case i see some traffic on eth0
   with tcpdump.

Ok, so at least it sounds like your TCP network is sanely configured.

   Now, when i installed and used Open MPI, things didn't work as
easy.. Here is
what happens. After recompiling the sources with the mpicc that comes with
   open-mpi:

   $ /usr/local/openmpi/bin/mpirun  --prefix /usr/local/openmpi --mca
   pls_rsh_agent ssh --mca btl tcp -np 2 --host 10.12.4.34,10.12.4.32
   /path/to/hello_world
   Hello, world, I am 0 of 2 and this is on : pe830-02.
   Hello, world, I am 1 of 2 and this is on: pe830-01.

So far so good, using eth0 interfaces.. hello_world works just fine. Now,
   when i try the broadcast program:

In reality, you always need to include two BTLs when specifying. You need both the one you want to use (mvapi,openib,tcp,etc.) and "self". You can run into issues otherwise.

   $ /usr/local/openmpi/bin/mpirun  --prefix /usr/local/openmpi --mca
   pls_rsh_agent ssh --mca btl tcp -np 2 --host 10.12.4.34,10.12.4.32
   /path/to/broadcast

It just hangs there, it doesn't prompt me the "Enter the vector length:" string . So i just enter a number anyway since i know the behavior of the
   program:

   10
   Enter the vector length: i am: 0 , and i have 5 vector elements
   i am: 1 , and i have 5 vector elements
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000
   [0] 10.000000

So, that's the first bump with the openmpi.. Now , if i try to use ib0
   interfaces instead of eth0 ones, i get:

I'm actually surprised this worked in LAM/MPI, to be honest. There should be an fflush() after the printf() to make sure that the output is actually sent out of the application.

   $  /usr/local/openmpi/bin/mpirun  --prefix /usr/local/openmpi --mca
pls_rsh_agent ssh --mca btl openib -np 2 --host 192.168.1.34,192.168.1.32
   /path/to/hello_world
---------------------------------------------------------------------- ----
   No available btl components were found!

This means that there are no components of this type installed on your
   system or all the components reported that they could not be used.

This is a fatal error; your MPI process is likely to abort. Check the output of the "ompi_info" command and ensure that components of this
   type are available on your system.  You may also wish to check the
value of the "component_path" MCA parameter and ensure that it has at
   least one directory that contains valid MCA components.

---------------------------------------------------------------------- ----
   [pe830-01.domain.com:05942]

I know, it thinks that it doesn't have openib components installed, however,
   ompi_info on both machines say otherwise:

   $ ompi_info | grep openib
   MCA mpool: openib (MCA v1.0, API v1.0, Component v1.0.2)
   MCA btl: openib (MCA v1.0, API v1.0, Component v1.0.2)

I don't think it will help, but can you try again with --mca btl openib,self? For some reason, it appears that the openib component is saying that it can't run.

   Now the questions are...
   1- In the case of using lam/mpi over ib0 interfaces.. Does lam/mpi
   automatically just use IPoIB ?

Yes, LAM has no idea what that Open IB thing is -- it just uses the ethernet device.

2 - Is there a tcpdump-like utility to dump the traffic on Infiniband HCAs?

I'm not aware of any, but that may occur.

3 - In the case of Open MPI, does --mca btl arg option have to be passed
   everytime? For example,

   $ /usr/local/openmpi/bin/mpirun  --prefix /usr/local/openmpi --mca
   pls_rsh_agent ssh --mca btl tcp -np 2 --host 10.12.4.34,10.12.4.32
   /path/to/hello_world

works just fine, but the same command without the "--mca btl tcp" bit gives
   the:

---------------------------------------------------------------------- ---- It looks like MPI_INIT failed for some reason; your parallel process is likely to abort. There are many reasons that a parallel process can fail during MPI_INIT; some of which are due to configuration or environment problems. This failure appears to be an internal failure; here's some
   additional information (which may only be relevant to an Open MPI
   developer):

     PML add procs failed
     --> Returned value -2 instead of OMPI_SUCCESS
---------------------------------------------------------------------- ----
   *** An error occurred in MPI_Init
   *** before MPI was initialized
   *** MPI_ERRORS_ARE_FATAL (goodbye)

   error ...

This makes it sound like Open IB is failing to setup properly. I'm a bit out of my league on this one -- is there any application you can run

   4 - How come the behavior of broadcast.c was different on Open MPI
than it is
   on lam/mpi?

I think I answered this one already.

5 - Any ideas as to why i am getting no btl component error when i want to use openib even though ompi_info shows it? If it help any further , I have
   the following openib modules :

This usually (but not always) indicates that something is going wrong with initializing the hardware interface. ompi_info only tries to load the module, but not initialize the network device.


Brian

--
  Brian Barrett
  Open MPI developer
  http://www.open-mpi.org/


Reply via email to