Rolf Vandevaart wrote:
I see the problem - really just a current limit on our implementation.
At the moment, our launchers don't know how to take advantage of
existing daemons on the remote nodes. Your mpirun is correctly
connecting to the persistent daemon on ct2, but will launch its own
daemons on the other nodes.
Thanks for your information. You said I could ask more so I am! See
Ralph Castain wrote On 03/30/06 16:51,:
I apologize for the scarce documentation - we are working on it, but
have a ways to go. I've tried to address your questions below. Please
feel free to ask more!
Rolf Vandevaart wrote:
I am new to the Open MPI world, and I have been trying to get a better
understanding of the ORTE environment. At this point, I have a few
questions that I was hoping someone could answer.
1. I have heard mention of running the ORTE daemons in persistent mode,
however, I can find no details of how to do this. Are there arguments
to either orted or mpirun to make this work right?
Normally, we start a persistent daemon with:
orted --seed --persistent --scope=public
This will start the daemon and "daemonize" it so it keeps running
until told to die. The arguments worth noting are:
(a) --persistent. Tells the daemon to "stay alive" until specifically
told to "die"
(b) --scope=[public, private, exclusive]. This actually pertains to
the universe, but you'll need to provide it anyway to ensure proper
connectivity to anything you try to run. Right now, the daemons
default to "exclusive", which means nothing can connect to them except
the application that spawned them - no value to anyone if started with
the above command! Private would exclude them to contact only from you
- I haven't tested this enough to guarantee its functionality. I
usually run them as "public" since security isn't a big concern right
now - all this means is that anyone who can read the session directory
tree (which is normally "locked" to only you anyway) would be able to
connect to the daemon.
(c) --seed. Indicates that this daemon is the first one and therefore
will host the data storage for the registry and other central services
(d) --universe=userid@hostname:universe_name. Allows you to name your
universe to whatever you like. We use this to allow you to have
multiple universes co-existing but separate - I've been explaining the
reasons for that elsewhere, but will send them to this list if
desired. You don't have to provide this, nor do you have to provide
all the fields (e.g., you could just say "--universe=foo" to set the
You can provide the same options to mpirun, if you like - mpirun will
simply start an orted and pass those parameters along, and the orted
will merrily stay alive after the specified application completes.
While I understand all that has been written here in theory, I am still
to get things to work.
The persistent daemon seems to be ignored when I do an mpirun. I have
system calls and looked at the process tree, and the persistent daemon
does not seem
to be part of the fun. So, I will be specific about what I am doing,
and maybe you can point
out what I am doing wrong.
I have a 3 node cluster. ct2, ct4, and ct5. I am launching the job
from ct2 and trying to
run on ct4 and ct5 which have persistent daemons on them. I have
selected the daemon
on ct4 to be the seed.
ct4> orted --seed --persistent --scope public -universe foo
ct5> orted --persistent --scope public -universe foo
ct2> mpirun --mca pls_rsh_agent rsh -np 4 -host ct4,ct5 -universe foo
While the program is running, I see this on ct4 and ct5.
ps -ef | grep orted
rolfv 9456 1 0 11:24:26 ? 0:00 orted --bootproxy 1
--name 0.0.2 --num_procs 3 --vpid_start 0 --nodename ct4
rolfv 9386 1 0 11:21:30 ? 0:00 orted --seed
--persistent --scope public --universe foo
Thanks for any additional details.
This is a known issue that we need to address - just low on the
priority list right now.
Tell you what I'll do - I'll implement an interface in orteconsole that
will let you execute the dump commands from there. The problem with
doing it from gdb is that gdb doesn't like the volume of text that can
result - so it keeps cutting off the dump output. Also, gdb -
particularly if you are using shared libraries instead of a static
build configuration - can bark when you try to access the registry
functions directly, depending upon when exactly you do so.
3. I have a similar question about orteprobe. Is this something
we should know about?
Yes and no - there's nothing secret about it. We use it internally to
OpenRTE to "probe" a machine and see if we have a daemon/universe
operating on it. Basically, we launch orteprobe on the remote machine
- it checks to see if a session directory exists on it, attempts to
connect to any universes it finds, and then reports back on its
findings. Based on that report, we either launch an orted on the
remote machine (to act as our surrogate so we can launch an
application on that cell) or connect to an existing universe on the
remote machine (and then tell it to launch the application for us).
4. Is there an easy way to view the data in the General Purpose
Registry? This may be related to my first question, in that I
could imagine having persistent daemons and then I would like
to see what is stored in the registry.
Well, yes and no. Ideally, that would be a command from within the
orteconsole function, but I don't think that has been implemented yet.
I'd be happy to do so, if that is something you would like (shouldn't
take long at all). There are a set of "dump" functions in the registry
API for just that purpose. I usually access them via gdb - I attach
the debugger to the orted process, then use the dump functions to
output the values in the registry.
What exactly do you type in for the dump functions? I saw these functions,
but could not get them to fire properly.
The orteconsole approach will be *much* cleaner. Only negative is that
you won't be able to look at the contents step-by-step as the
If you really want to do it from gdb, then you have to type in the name
of the replica's dump functions (e.g., orte_gpr_replica_dump_all) or
the proxy's dump functions (e.g., orte_gpr_proxy_dump_all) directly -
you can't go through the orte_gpr.xxx interface. Obviously, you use the
replica commands on the persistent daemon (you'll probably want to tell
it --no-daemonize so gdb can remain connected to it) and the proxy
commands if you are working on one of the application processes.
Doing this will allow you to step through the process launch
progression within mpirun (or the persistent daemon, if you are using
it) and see what info is being stored where/when. Like I said, though,
gdb will truncate the output, so you won't see everything :-(