Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] callback debugging
From: Adrian Reber (adrian_at_[hidden])
Date: 2014-01-21 14:56:49


Good to know that it does not make any sense. So it not just me.

Looking at the call chain I can see

orte_snapc_base_select(ORTE_PROC_IS_HNP, !ORTE_PROC_IS_DAEMON);

and the second parameter is used to decide if it is an app or not:

int orte_snapc_base_select(bool seed, bool app) in orte/mca/snapc/base/snapc_base_select.c

and if it is true the code with the barrier is used.

In orte/mca/snapc/base/snapc_base_select.c there is also following
comment:

/* XXX -- TODO -- framework_subsytem -- this shouldn't be necessary once the framework system is in place */

Is this something which needs to be changed and which might be the cause
for this problem?

On Tue, Jan 21, 2014 at 07:27:32AM -0800, Ralph Castain wrote:
> That doesn't make any sense - I can't imagine a reason for orte-checkpoint itself to be running a barrier. I wonder if it is selecting the wrong component in snapc?
>
> As for the patch, that isn't going to work. The collective id has to be *globally* unique, which means that only orterun can issue a new one. So you have to get thru orte_init before you can request one as it requires a communication.
>
> However, like I said, it makes no sense for orte-checkpoint to do a barrier as it is a singleton - there is nothing for it to "barrier" with.
>
> On Jan 21, 2014, at 7:24 AM, Adrian Reber <adrian_at_[hidden]> wrote:
>
> > I think I still do not really understand how it works.
> >
> > The barrier on which orte-checkpoint is currently hanging is in
> > app_coord_init(). You are also saying that orte-checkpoint
> > should not be calling a barrier. The backtrace of the point where it
> > is hanging now looks like:
> >
> > #0 0x00007ffff69befa0 in __nanosleep_nocancel () at ../sysdeps/unix/syscall-template.S:81
> > #1 0x00007ffff7b45712 in app_coord_init () at ../../../../../orte/mca/snapc/full/snapc_full_app.c:208
> > #2 0x00007ffff7b3a5ce in orte_snapc_full_module_init (seed=false, app=true) at ../../../../../orte/mca/snapc/full/snapc_full_module.c:207
> > #3 0x00007ffff7b375de in orte_snapc_base_select (seed=false, app=true) at ../../../../orte/mca/snapc/base/snapc_base_select.c:96
> > #4 0x00007ffff7a9884a in orte_ess_base_tool_setup () at ../../../../orte/mca/ess/base/ess_base_std_tool.c:192
> > #5 0x00007ffff7a9fe85 in rte_init () at ../../../../../orte/mca/ess/tool/ess_tool_module.c:83
> > #6 0x00007ffff7a4647f in orte_init (pargc=0x7fffffffd94c, pargv=0x7fffffffd940, flags=8) at ../../orte/runtime/orte_init.c:158
> > #7 0x0000000000402859 in ckpt_init (argc=51, argv=0x7fffffffda78) at ../../../../orte/tools/orte-checkpoint/orte-checkpoint.c:610
> > #8 0x0000000000401d7a in main (argc=51, argv=0x7fffffffda78) at ../../../../orte/tools/orte-checkpoint/orte-checkpoint.c:245
> >
> > Maybe I am doing something completely wrong. I am currently
> > running 'orterun -np 2 test-programm'.
> >
> > In another terminal I am starting orte-checkpoint with the PID of
> > orterun and the barrier in app_coord_init() is just before it tries
> > to communicate with orterun. Is this the correct setup?
> >
> > Adrian
> >
> > On Mon, Jan 20, 2014 at 05:33:59PM -0600, Josh Hursey wrote:
> >> If it is the application, then there is probably a barrier in the
> >> app_coord_init() to make sure all the applications are up and running.
> >> After this point then the global coordinator knows that the application can
> >> be checkpointed.
> >>
> >> I don't think orte-checkpoint should be calling a barrier - from what I
> >> recall.
> >>
> >>
> >> On Mon, Jan 20, 2014 at 4:46 PM, Ralph Castain <rhc_at_[hidden]> wrote:
> >>
> >>> Is it orte-checkpoint that is hanging, or the app you are trying to
> >>> checkpoint?
> >>>
> >>>
> >>> On Jan 20, 2014, at 2:10 PM, Adrian Reber <adrian_at_[hidden]> wrote:
> >>>
> >>> Thanks for your help. I tried initializing the barrier correctly (see
> >>> attached patch) but now, instead of crashing, it just hangs on the
> >>> barrier while running orte-checkpoint
> >>>
> >>> [dcbz:20150] [[41665,0],0] grpcomm:bad entering barrier
> >>> [dcbz:20150] [[41665,0],0] ACTIVATING GRCPCOMM OP 0 at
> >>> ../../../../../orte/mca/grpcomm/bad/grpcomm_bad_module.c:206
> >>>
> >>> #0 0x00007ffff69befa0 in __nanosleep_nocancel () at
> >>> ../sysdeps/unix/syscall-template.S:81
> >>> #1 0x00007ffff7b456ba in app_coord_init () at
> >>> ../../../../../orte/mca/snapc/full/snapc_full_app.c:207
> >>> #2 0x00007ffff7b3a582 in orte_snapc_full_module_init (seed=false,
> >>> app=true) at ../../../../../orte/mca/snapc/full/snapc_full_module.c:207
> >>>
> >>> it hangs looping at ORTE_WAIT_FOR_COMPLETION(coll->active);
> >>>
> >>> I do not understand on what the barrier here is actually waiting for. Where
> >>> do I need to look to find the place the barrier is waiting for?
> >>>
> >>> I also tried initializing the collective id's in
> >>> orte/mca/plm/base/plm_base_launch_support.c but that code is never
> >>> used running the orte-checkpoint tool
> >>>
> >>> Adrian
> >>>
> >>> On Sat, Jan 11, 2014 at 11:46:42AM -0800, Ralph Castain wrote:
> >>>
> >>> I took a look at this, and I'm afraid you have some work to do in the
> >>> orte/mca/snapc code base:
> >>>
> >>> 1. you must use dynamically allocated buffers for rml.send_buffer_nb. See
> >>> r30261 for an example of the changes that need to be made - I did some, but
> >>> can't swear to catching them all. It was enough to at least get a proc past
> >>> the initial snapc registration
> >>>
> >>> 2. you are reusing collective id's to execute several orte_grpcomm.barrier
> >>> calls - those ids are used elsewhere during MPI_Init. This is not allowed -
> >>> a collective id can only be used *once*. What you need to do is go into
> >>> orte/mca/plm/base/plm_base_launch_support.c and (when cr is configured) add
> >>> cr-specific collective id's for this purpose. I don't know how many places
> >>> in the cr code create their own barriers, but they each need a collective
> >>> id.
> >>>
> >>> If you prefer and have the time, you are welcome to extend the collective
> >>> code to allow id reuse. This would require that each daemon and app "reset"
> >>> the collective fields when a collective is declared complete. It isn't that
> >>> hard to do - just never had a reason to do it. I can take a shot at it when
> >>> time permits (may have some time this weekend)
> >>>
> >>> 3. when you post the non-blocking recv in the snapc/full code, it looks to
> >>> me like you need to block until you get the answer. I don't know where in
> >>> the code flow this is occurring - if you are not in an event, then it is
> >>> okay to block using ORTE_WAIT_FOR_COMPLETION. Look in
> >>> orte/mca/routed/base/routed_base_fns.c starting at line 252 for an example.
> >>>
> >>> HTH
> >>> Ralph
> >>>
> >>> On Jan 10, 2014, at 12:55 PM, Ralph Castain <rhc_at_[hidden]> wrote:
> >>>
> >>>
> >>> On Jan 10, 2014, at 12:45 PM, Adrian Reber <adrian_at_[hidden]> wrote:
> >>>
> >>> On Fri, Jan 10, 2014 at 09:48:14AM -0800, Ralph Castain wrote:
> >>>
> >>>
> >>> On Jan 10, 2014, at 8:02 AM, Adrian Reber <adrian_at_[hidden]> wrote:
> >>>
> >>> I am currently trying to understand how callbacks are working. Right now
> >>> I am looking at orte/mca/rml/base/rml_base_receive.c
> >>> orte_rml_base_comm_start() which does
> >>>
> >>> orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD,
> >>> ORTE_RML_TAG_RML_INFO_UPDATE,
> >>> ORTE_RML_PERSISTENT,
> >>> orte_rml_base_recv,
> >>> NULL);
> >>>
> >>> As far as I understand it orte_rml_base_recv() is the callback function.
> >>> At which point should this function run? When the data is actually
> >>> received?
> >>>
> >>>
> >>> Not precisely. When data is received by the OOB, it pushes the data into
> >>> an event. When that event gets serviced, it calls the orte_rml_base_receive
> >>> function which processes the data to find the matching tag, and then uses
> >>> that to execute the callback to the user code.
> >>>
> >>>
> >>> The same for send_buffer_nb() functions. I do not see the callback
> >>> functions actually running. How can I verify that the callback functions
> >>> are running. Especially for the send case it sounds pretty obvious how
> >>> it should work but I never see the callback function running. At least
> >>> in my setup.
> >>>
> >>>
> >>> The data is not immediately sent. It gets pushed into an event. When that
> >>> event gets serviced, it calls the orte_oob_base_send function which then
> >>> passes the data to each active OOB component until one of them says it can
> >>> send it. The data is then pushed into another event to get it into the
> >>> event base for that component's active module - when that event gets
> >>> serviced, the data is sent. Once the data is sent, an event is created
> >>> that, when serviced, executes the callback to the user code.
> >>>
> >>> If you aren't seeing callbacks, the most likely cause is that the orte
> >>> progress thread isn't running. Without it, none of this will work.
> >>>
> >>>
> >>> Thanks. Running configure without '--with-ft=cr' I can run a program and
> >>> use orte-top. In orterun I can see that the callback is running and
> >>> orte-top displays the retrieved information. I can also see in orte-top
> >>> that the callbacks are working.
> >>>
> >>>
> >>> Actually, I'm rather impressed - I hadn't tested orte-top and didn't
> >>> honestly know if it would work any more! Glad to hear it does :-)
> >>>
> >>> Doing the same with '--with-ft=cr'
> >>> enabled orte-top crashes as well as orte-checkpoint and both (-top and
> >>> -checkpoint) seem to no longer have working callbacks and that is why
> >>> they are probably crashing. So some code which is enabled by '--with-ft=cr'
> >>> seems to break callbacks in orte-top as well as in orte-checkpoint.
> >>> orterun handles callbacks no matter if configured with or without
> >>> '--with-ft=cr'.
> >>>
> >>>
> >>> I can take a look this weekend - probably something silly
> >>>
> >>>
> >>> Adrian
> >>>
> >>> <grpcomm.txt>_______________________________________________
> >>>
> >>> devel mailing list
> >>> devel_at_[hidden]
> >>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
> >>>
> >>>
> >>>
> >>> _______________________________________________
> >>> devel mailing list
> >>> devel_at_[hidden]
> >>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
> >>>
> >>
> >>
> >>
> >> --
> >> Joshua Hursey
> >> Assistant Professor of Computer Science
> >> University of Wisconsin-La Crosse
> >> http://cs.uwlax.edu/~jjhursey
> >
> >> _______________________________________________
> >> devel mailing list
> >> devel_at_[hidden]
> >> http://www.open-mpi.org/mailman/listinfo.cgi/devel
> >
> >
> > Adrian
> >
> > --
> > Adrian Reber <adrian_at_[hidden]> http://lisas.de/~adrian/
> > QOTD:
> > "I tried buying a goat instead of a lawn tractor; had to return
> > it though. Couldn't figure out a way to connect the snow blower."
> > _______________________________________________
> > devel mailing list
> > devel_at_[hidden]
> > http://www.open-mpi.org/mailman/listinfo.cgi/devel
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel

                Adrian

-- 
Adrian Reber <adrian_at_[hidden]>            http://lisas.de/~adrian/
Hempstone's Question:
	If you have to travel on the Titanic, why not go first class?