Hey all,
Sorry for my lag on this thread, I'm still settling back into
Bloomington and catching up on email traffic.
This is certainly my fault WRT the addition of the RDS call to
orte_init_stage1(). I never tested the case where a process is a
singleton and not the seed. :(
Since the RAS (or functionality represented by this subsystem) was
exposed at this level, it was assumed that the RDS is also active at
this time. The addition in orte_init_stage1 was to add host entries to
both the RAS and RDS (instead of just the RAS) when we start a
singleton process.
A quick repair would be to protect the RDS section from all non-seed
processes. E.g.
if(orte_process_info.seed) {
ret = orte_rds.store_resource(&rds_single_host);
if (ORTE_SUCCESS != ret ) {
ORTE_ERROR_LOG(ret);
return ret;
}
}
An additional fix would be to add a call to the rmgr to setup
singleton
processes, thus pulling out the 'singleton process only' chunk of code
from the orte_init_stage1() and into the rmgr. Something like:
if (orte_process_info.singleton) {
if (ORTE_SUCCESS != (ret =
orte_rmgr_base_setup_singleton(my_jobid,...))) {
ORTE_ERROR_LOG(ret);
return ret;
}
}
Currently this would only contain the addition of the singleton
process
to the RDS and RAS, but Ralph mentioned last week that he ran across
some other 'singleton only' stuff that might be needed.
Is there a design issue in adding this functionality to the rmgr, with
the proper protection around access to the RDS?
I guess my overall argument is that the RDS should be called in the
singleton+seed case since we are adding resources to the allocation
[RAS], and thus the resources globally available [RDS]. Do we assume
that if the process is a singleton and not the seed then it has
already
been placed in the RDS, and only needs to confirm it allocation in the
RAS? Shouldn't that registry handling only happen at the seed level if
we assume it has launched the singleton process?
It is likely that I could have things confused a bit with how we
define
a singleton process, and how they are created with relation to the
seed.
As a general bug notice in ORTE: There is an outstanding bug in the
proxy/replica NS components when creating new cellid's that I ran
across last Friday, before I had to stop. Something is getting mangled
in the packing of the command sent to the seed. I had to wrap up
before
I could seek a good fix, just enough to characterize the problem.
Thoughts?
Sorry for causing trouble,
Josh
On Aug 18, 2005, at 3:33 PM, Tim S. Woodall wrote:
I'm seeing a problem in orte_init_stage1 when running w/ a persistent
daemon.
The problem is that the orte_inti call attempts to call rds subsystem
directly,
which is not supposed to be exposed at that level. rds is used
internally by
the rmgr - and only initialized on the seed. The proxy rmgr is loaded
when
a persistent daemon is available - and therefore the rds is not
loaded.
So... orte_init_stage1 shouldn't be calling rds directly...
Tim
Brian Barrett wrote:
Yeah, although there really shouldn't be a way for the pointer to be
NULL. Was this a static build? I was seeing some weird memory
issues on static builds last night... I'll take a look on odin and
see what I can find.
Brian
On Aug 18, 2005, at 11:18 AM, Tim S. Woodall wrote:
Brian,
Wasn't the introduction of sds part of your changes for redstorm?
Any ideas
why it would be NULL here?
Thanks,
Tim
Rainer Keller wrote:
Hello,
see the "same" (well probably not exactly same) thing here in
Opteron with
64bit (-g and so on), I get:
#0 0x0000000040085160 in orte_sds_base_contact_universe ()
at ../../../../../orte/mca/sds/base/sds_base_interface.c:29
29 return orte_sds_base_module->contact_universe();
(gdb) where
#0 0x0000000040085160 in orte_sds_base_contact_universe ()
at ../../../../../orte/mca/sds/base/sds_base_interface.c:29
#1 0x0000000040063e95 in orte_init_stage1 ()
at ../../../orte/runtime/orte_init_stage1.c:185
#2 0x0000000040017e7d in orte_system_init ()
at ../../../orte/runtime/orte_system_init.c:38
#3 0x00000000400148f5 in orte_init () at ../../../orte/runtime/
orte_init.c:46
#4 0x000000004000dfc7 in main (argc=4, argv=0x7fbfffe8a8)
at ../../../../orte/tools/orterun/orterun.c:291
#5 0x0000002a95c0c017 in __libc_start_main () from
/lib64/libc.so.6
#6 0x000000004000bf2a in _start ()
(gdb)
within mpirun
orte_sds_base_module here is Null...
This is without persistent orted; Just mpirun...
CU,
ray
On Thursday 18 August 2005 16:57, Nathan DeBardeleben wrote:
FYI, this only happens when I let OMPI compile 64bit on Linux.
When I
throw in there CFLAGS=FFLAGS=CXXFLAGS=-m32 orted, my myriad of
test
codes, mpirun, registry subscription codes, and JNI all work like
a champ.
Something's wrong with the 64bit it appears to me.
-- Nathan
Correspondence
-----------------------------------------------------------------
--
-
-
Nathan DeBardeleben, Ph.D.
Los Alamos National Laboratory
Parallel Tools Team
High Performance Computing Environments
phone: 505-667-3428
email: ndeb...@lanl.gov
-----------------------------------------------------------------
--
-
-
Tim S. Woodall wrote:
Nathan,
I'll try to reproduce this sometime this week - but I'm pretty
swamped.
Is Greg also seeing the same behavior?
Thanks,
Tim
Nathan DeBardeleben wrote:
To expand on this further, orte_init() seg faults on both
bluesteel
(32bit linux) and sparkplug (64bit linux) equally. The
required
condition is that orted must be running first (which of course
we
require for our work - a persistent orte daemon and registry).
[bluesteel]~/ptp > ./dump_info
Segmentation fault
[bluesteel]~/ptp > gdb dump_info
GNU gdb 6.1
Copyright 2004 Free Software Foundation, Inc.
GDB is free software, covered by the GNU General Public
License, and
you are
welcome to change it and/or distribute copies of it under
certain
conditions.
Type "show copying" to see the conditions.
There is absolutely no warranty for GDB. Type "show warranty"
for
details.
This GDB was configured as "x86_64-suse-linux"...Using host
libthread_db library "/lib64/tls/libthread_db.so.1".
(gdb) run
Starting program: /home/ndebard/ptp/dump_info
Program received signal SIGSEGV, Segmentation fault.
0x0000000000000000 in ?? ()
(gdb) where
#0 0x0000000000000000 in ?? ()
#1 0x000000000045997d in orte_init_stage1 () at
orte_init_stage1.c:419
#2 0x00000000004156a7 in orte_system_init () at
orte_system_init.c:38
#3 0x00000000004151c7 in orte_init () at orte_init.c:46
#4 0x0000000000414cbb in main (argc=1, argv=0x7fbffff298) at
dump_info.c:185
(gdb)
-- Nathan
Correspondence
---------------------------------------------------------------
--
-
---
Nathan DeBardeleben, Ph.D.
Los Alamos National Laboratory
Parallel Tools Team
High Performance Computing Environments
phone: 505-667-3428
email: ndeb...@lanl.gov
---------------------------------------------------------------
--
-
---
Nathan DeBardeleben wrote:
Just to clarify:
1: no orted started (meaning the MPIrun or registry programs
will
start one by themselves) causes those programs to lock up.
2: starting orted by hand (trying to get these programs to
connect to
a centralized one) causes the connecting programs to seg
fault.
-- Nathan
Correspondence
--------------------------------------------------------------
--
-
----
Nathan DeBardeleben, Ph.D.
Los Alamos National Laboratory
Parallel Tools Team
High Performance Computing Environments
phone: 505-667-3428
email: ndeb...@lanl.gov
--------------------------------------------------------------
--
-
----
Nathan DeBardeleben wrote:
So I dropped an .ompi_ignore into that directory,
reconfigured, and
compile worked (yay!).
However, not a lot of progress: mpirun locks up, all my
registry test
programs lock up as well. If I start the orted by hand, then
any of my
registry calling programs cause segfault:
[sparkplug]~/ptp > gdb sub_test
GNU gdb 6.1
Copyright 2004 Free Software Foundation, Inc.
GDB is free software, covered by the GNU General Public
License, and
you are
welcome to change it and/or distribute copies of it under
certain
conditions.
Type "show copying" to see the conditions.
There is absolutely no warranty for GDB. Type "show
warranty" for
details.
This GDB was configured as "x86_64-suse-linux"...Using host
libthread_db library "/lib64/tls/libthread_db.so.1".
(gdb) run
Starting program: /home/ndebard/ptp/sub_test
Program received signal SIGSEGV, Segmentation fault.
0x0000000000000000 in ?? ()
(gdb) where
#0 0x0000000000000000 in ?? ()
#1 0x00000000004598a5 in orte_init_stage1 () at
orte_init_stage1.c:419 #2 0x00000000004155cf in
orte_system_init ()
at orte_system_init.c:38 #3 0x00000000004150ef in orte_init
() at
orte_init.c:46
#4 0x00000000004148a1 in main (argc=1, argv=0x7fbffff178)
at
sub_test.c:60
(gdb)
Yes, I recompiled everything.
Here's an example of me trying something a little more
complicated
(which I believe locks up for the same reason - something
borked with
the registry interaction).
[sparkplug]~/ompi-test > bjssub -s 10000 -n 10 -i bash
Waiting for interactive job nodes.
(nodes 18 16 17 18 19 20 21 22 23 24 25)
Starting interactive job.
NODES=16,17,18,19,20,21,22,23,24,25
JOBID=18
so i got my nodes
ndebard@sparkplug:~/ompi-test> export
OMPI_MCA_ptl_base_exclude=sm
ndebard@sparkplug:~/ompi-test> export
OMPI_MCA_pls_bproc_seed_priority=101
and set these envvars like we need to use Greg's bproc,
without the
2nd export the machine's load maxes and locks up.
ndebard@sparkplug:~/ompi-test> bpstat
Node(s) Status Mode
User Group 100-128 down
---------- root root 0-15
up ---x------ vchandu vchandu
16-25 up
---x------
ndebard ndebard
26-27 up
---x------
root root 28-30 up
---x--x--x root root ndebard@sparkplug:~/ompi-test>
env | grep
NODES
NODES=16,17,18,19,20,21,22,23,24,25
yes, i really have the nodes
ndebard@sparkplug:~/ompi-test> mpicc -o test-mpi test-mpi.c
ndebard@sparkplug:~/ompi-test>
recompile for good measure
ndebard@sparkplug:~/ompi-test> ls /tmp/openmpi-sessions-
ndebard*
/bin/ls: /tmp/openmpi-sessions-ndebard*: No such file or
directory
proof that there's no left over old directory
ndebard@sparkplug:~/ompi-test> mpirun -np 1 test-mpi
it never responds at this point - but I can kill it with ^C.
mpirun: killing job...
Killed
ndebard@sparkplug:~/ompi-test>
-- Nathan
Correspondence
-------------------------------------------------------------
--
-
-----
Nathan DeBardeleben, Ph.D.
Los Alamos National Laboratory
Parallel Tools Team
High Performance Computing Environments
phone: 505-667-3428
email: ndeb...@lanl.gov
-------------------------------------------------------------
--
-
-----
Jeff Squyres wrote:
Is this what Tim Prins was working on?
On Aug 16, 2005, at 5:21 PM, Tim S. Woodall wrote:
I'm not sure why this is even building... Is someone
working on this?
I thought we had .ompi_ignore files in this directory.
Tim
Nathan DeBardeleben wrote:
So I'm seeing all these nice emails about people
developing on OMPI
today yet I can't get it to compile. Am I out here in
limbo on this
or
are others in the same boat? The errors I'm seeing are
about some
bproc
code calling undefined functions and they are linked again
below.
-- Nathan
Correspondence
----------------------------------------------------------
--
-
-------
- Nathan DeBardeleben, Ph.D.
Los Alamos National Laboratory
Parallel Tools Team
High Performance Computing Environments
phone: 505-667-3428
email: ndeb...@lanl.gov
----------------------------------------------------------
--
-
-------
-
Nathan DeBardeleben wrote:
Back from training and trying to test this but now OMPI
doesn't
compile
at all:
gcc -DHAVE_CONFIG_H -I. -I. -I../../../../include
-I../../../../include -I../../../.. -I../../../..
-I../../../../include -I../../../../opal
-I../../../../orte
-I../../../../ompi -g -Wall -Wundef -Wno-long-long -
Wsign-compare
-Wmissing-prototypes -Wstrict-prototypes -Wcomment -
pedantic
-Werror-implicit-function-declaration -fno-strict-
aliasing -MT
ras_lsf_bproc.lo -MD -MP -MF .deps/ras_lsf_bproc.Tpo -c
ras_lsf_bproc.c -o ras_lsf_bproc.o
ras_lsf_bproc.c: In function
`orte_ras_lsf_bproc_node_insert':
ras_lsf_bproc.c:32: error: implicit declaration of
function
`orte_ras_base_node_insert'
ras_lsf_bproc.c: In function
`orte_ras_lsf_bproc_node_query':
ras_lsf_bproc.c:37: error: implicit declaration of
function
`orte_ras_base_node_query'
make[4]: *** [ras_lsf_bproc.lo] Error 1
make[4]: Leaving directory
`/home/ndebard/ompi/orte/mca/ras/lsf_bproc'
make[3]: *** [all-recursive] Error 1
make[3]: Leaving directory `/home/ndebard/ompi/orte/mca/
ras'
make[2]: *** [all-recursive] Error 1
make[2]: Leaving directory `/home/ndebard/ompi/orte/mca'
make[1]: *** [all-recursive] Error 1
make[1]: Leaving directory `/home/ndebard/ompi/orte'
make: *** [all-recursive] Error 1
[sparkplug]~/ompi >
Clean SVN checkout this morning with configure:
[sparkplug]~/ompi > ./configure --enable-static --
disable-shared
--without-threads --prefix=/home/ndebard/local/ompi
--with-devel-headers
-- Nathan
Correspondence
---------------------------------------------------------
--
-
-------
-- Nathan DeBardeleben, Ph.D.
Los Alamos National Laboratory
Parallel Tools Team
High Performance Computing Environments
phone: 505-667-3428
email: ndeb...@lanl.gov
---------------------------------------------------------
--
-
-------
--
Brian Barrett wrote:
This is now fixed in SVN. You should no longer need the
--build=i586... hack to compile 32 bit code on
Opterons.
Brian
On Aug 12, 2005, at 3:17 PM, Brian Barrett wrote:
On Aug 12, 2005, at 3:13 PM, Nathan DeBardeleben wrote:
We've got a 64bit Linux (SUSE) box here. For a
variety of
reasons (Java, JNI, linking in with OMPI libraries,
etc which I
won't get into)
I need to compile OMPI 32 bit (or get 64bit versions
of a lot of
other
libraries).
I get various compile errors when I try different
things, but
first
let
me explain the system we have:
<snip>
This goes on and on and on actually. And the 'is
incompatible
with
i386:x86-64 output' looks to be repeated for every
line before
this
error which actually caused the Make to bomb.
Any suggestions at all? Surely someone must have
tried to force
OMPI to
build in 32bit mode on a 64bit machine.
I don't think anyone has tried to build 32 bit on an
Opteron,
which is the cause of the problems...
I think I know how to fix this, but won't happen until
later in
the weekend. I can't think of a good workaround until
then.
Well, one possibility is to set the target like you
were doing
and disable ROMIO. Actually, you'll also need to
disable
Fortran 77. So something like:
./configure [usual options] --build=i586-suse-linux --
disable-io-
romio --disable-f77
might just do the trick.
Brian
--
Brian Barrett
Open MPI developer
http://www.open-mpi.org/
----
Josh Hursey
jjhur...@open-mpi.org
http://www.open-mpi.org/
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
