Am 28.02.2014 um 21:23 schrieb Brice Goglin: > OK, the problem is that node14's BIOS reports invalid NUMA info. It properly > detects 2 sockets with 16-cores each. But it reports 2 NUMA nodes total, > instead of 2 per socket (4 total). And hwloc warns because the cores > contained in these NUMA nodes are incompatible with sockets: > socket0 contains 0-15 > socket1 contains 16-23 > NUMA node0 contains 0-7+16-23 > NUMA node1 contains 8-15+24-31 > > > Interestingly the BIOS is more recent on the broken node: > --- node15/sys/class/dmi/id/bios_date 2014-02-27 22:26:57.000000000 +0100 > +++ node14/sys/class/dmi/id/bios_date 2014-02-28 01:16:16.000000000 +0100 > @@ -1 +1 @@ > -08/31/2012 > +11/25/2013 > --- node15/sys/class/dmi/id/bios_version 2014-02-27 22:26:57.000000000 > +0100 > +++ node14/sys/class/dmi/id/bios_version 2014-02-28 01:16:16.000000000 > +0100 > @@ -1 +1 @@ > -3.0 > +3.5 > > > Unfortunately, 3.5 is the latest BIOS available for this machine on > supermicro's website, and 3.0 isn't available online afaict. You should > contact supermicro to report the bug and ask them to provide the old 3.0 BIOS > in the meantime.
One thing to try: load default values in the BIOS. Sometimes they get screwed up. Then: some flash applications allow to save the original BIOS before any upgrade. Maybe you can save the BIOS of one of the old nodes this way and install it on the new node. Disclaimer: do on your risk of course. -- Reuti > You could try running "lstopo foo.xml" on node15, make foo.xml available on > node14, and pass HWLOC_XMLFILE=/path/to/foo.xml in your environment. May need > some tweaks, but could work. > > Brice > > > > Le 28/02/2014 20:55, Gus Correa a écrit : >> Hi Brice and Ralph >> >> Many thanks for helping out with this! >> >> Yes, you are right about node15 being OK. >> Node15 was a red herring, as along with node14 it was part of >> the same job that failed. >> However, after a closer look, I noticed that failure reported >> by hwloc was indeed in node14. >> >> I attach both diagnostic files generated by hwloc-gather-topology on >> node14. >> >> I will open the node and see if there is anything unusual with the >> hardware, and perhaps reinstall the OS, as Ralph suggested. >> It is awkward that the other node that had the motherboard replaced >> passes the hwloc-gather-topology test. >> After motherboard replacement I renistalled the OS on both, >> but it doesn't hurt to do it again. >> >> Gus Correa >> >> >> >> >> On 02/28/2014 03:26 AM, Brice Goglin wrote: >>> Hello Gus, >>> I'll need the tarball generated by gather-topology on node14 to debug >>> this. node15 doesn't have any issue. >>> We've seen issues on AMD machines because of buggy BIOS reporting >>> incompatible Socket and NUMA info. If node14 doesn't have the same BIOS >>> version as other nodes, that could explain things. >>> Brice >>> >>> >>> >>> >>> Le 28/02/2014 01:39, Gus Correa a écrit : >>>> Thank you, Ralph! >>>> >>>> I did a bit more of homework, and found out that all jobs that had >>>> the hwloc error involved one specific node (node14). >>>> >>>> The "report bindings" output in those jobs' stderr show >>>> that node14 systematically failed to bind the processes to the cores, >>>> while other nodes on the same jobs didn't fail. >>>> Interestingly, the jobs continued to run, although they >>>> eventually failed, but much later. >>>> So, the hwloc error doesn't seem to stop the job on its tracks. >>>> As a matter of policy, should it perhaps shutdown the job instead? >>>> >>>> In addition, when I try the hwloc-gather-topology diagnostic on node14 >>>> I get the same error, a bit more verbose (see below). >>>> So, now my guess is that this may be a hardware problem on that node. >>>> >>>> I replaced two nodes' motherboards last week, including node14's, >>>> and something may have gone wrong on that one. >>>> The other node that had the motherboard replaced >>>> doesn't show the hwloc-gather-topology error, though. >>>> >>>> Does the error message below (Socket P#0 ...) >>>> suggest anything that I should be looking for on the hardware side? >>>> (Thermal compound on the heatsink, memory modules, etc) >>>> >>>> Thank you, >>>> Gus Correa >>>> >>>> >>>> >>>> [root@node14 ~]# /usr/bin/hwloc-gather-topology /tmp/$(uname -n) >>>> Hierarchy gathered in /tmp/node14.tar.bz2 and kept in >>>> /tmp/tmp.D46Sdhcnru/node14/ >>>> **************************************************************************** >>>> >>>> >>>> * Hwloc has encountered what looks like an error from the operating >>>> system. >>>> * >>>> * object (Socket P#0 cpuset 0x0000ffff) intersection without inclusion! >>>> * Error occurred in topology.c line 718 >>>> * >>>> * Please report this error message to the hwloc user's mailing list, >>>> * along with the output from the hwloc-gather-topology.sh script. >>>> **************************************************************************** >>>> >>>> >>>> Expected topology output stored in /tmp/node14.output >>>> >>>> >>>> On 02/27/2014 06:39 PM, Ralph Castain wrote: >>>>> The hwloc in 1.6.5 is very old (v1.3.2), so it's possible it is having >>>> trouble with those data/instruction cache breakdowns. >>>> I don't know why it wouldn't have shown up before, >>>> however, as this looks to be happening when we first try to >>>> assemble the topology. To check that, what happens if you just run >>>> "mpiexec hostname" on the local node? >>>>> >>>>> >>>>> On Feb 27, 2014, at 3:04 PM, Gus Correa<g...@ldeo.columbia.edu> wrote: >>>>> >>>>>> Dear OMPI pros >>>>>> >>>>>> This seems to be a question in the nowhere land between OMPI and hwloc. >>>>>> However, it appeared as an OMPI error, hence it may be OK to ask the >>>>>> question in this list. >>>>>> >>>>>> *** >>>>>> >>>>>> A user here got this error (or warning?) message today: >>>>>> >>>>>> + mpiexec -np 64 $HOME/echam-aiv_ldeo_6.1.00p1/bin/echam6 >>>>>> **************************************************************************** >>>>>> >>>>>> >>>>>> * Hwloc has encountered what looks like an error from the operating >>>>>> system. >>>>>> * >>>>>> * object intersection without inclusion! >>>>>> * Error occurred in topology.c line 594 >>>>>> * >>>>>> * Please report this error message to the hwloc user's mailing list, >>>>>> * along with the output from the hwloc-gather-topology.sh script. >>>>>> **************************************************************************** >>>>>> >>>>>> >>>>>> >>>>>> Additional info: >>>>>> >>>>>> 1) We have OMPI 1.6.5. This user is using the one built >>>>>> with Intel compilers 2011.13.367. >>>>>> >>>>>> 2) I set these MCA parameters in $OMPI/etc/openmpi-mca-params.conf >>>>>> (includes binding to core): >>>>>> >>>>>> btl = ^tcp >>>>>> orte_tag_output = 1 >>>>>> rmaps_base_schedule_policy = core >>>>>> orte_process_binding = core >>>>>> orte_report_bindings = 1 >>>>>> opal_paffinity_alone = 1 >>>>>> >>>>>> >>>>>> 3) The machines have dual-socket 16-core AMD Opteron 6376 (Abu-Dhabi), >>>>>> which have one FPU for each pair of cores, a hierarchy of caches >>>>>> serving >>>>>> sub-groups of cores, etc. >>>>>> The OS is Linux CentOS 6.4 with stock CentOS OFED. >>>>>> Interconnect is Infiniband QDR (Mellanox HW). >>>>>> >>>>>> 4) We have Torque 4.2.5, built with cpuset support. >>>>>> OMPI is built with Torque (tm) support. >>>>>> >>>>>> 5) In case it helps, I attach the output of >>>>>> hwloc-gather-topology, which I ran on the node that threw the error, >>>>>> although not immediately after the job failure. >>>>>> I used the hwloc-gather-topology script that comes with >>>>>> the hwloc (version 1.5) provided by CentOS. >>>>>> As far as I can tell the hwloc nuts and bits built into OMPI >>>>>> do not include the hwloc-gather-topology script (although it may be >>>>>> a newer hwloc version. 1.8 perhaps?). >>>>>> Hopefully the mail servers won't chop off the attachments. >>>>>> >>>>>> 6) I am a bit surprised by this error message, because I haven't >>>>>> seen it before, although we have used OMPI 1.6.5 in >>>>>> this machine with several other programs without problems. >>>>>> Alas, it happened now. >>>>>> >>>>>> ** >>>>>> >>>>>> - Is this a known hwloc problem in this processor architecture? >>>>>> >>>>>> - Is this a known issue in this combination of HW and SW? >>>>>> >>>>>> - Would not binding the MPI processes (to core or socket), perhaps >>>>>> help? >>>>>> >>>>>> - Any workarounds or suggestions? >>>>>> >>>>>> ** >>>>>> >>>>>> Thank you, >>>>>> Gus Correa >>>>>> <node15.output><node15.tar.bz2>_______________________________________________ >>>>>> >>>>>> >>>>>> users mailing list >>>>>> us...@open-mpi.org >>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users >>>>> >>>>> _______________________________________________ >>>>> users mailing list >>>>> us...@open-mpi.org >>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users >>>> >>>> _______________________________________________ >>>> users mailing list >>>> us...@open-mpi.org >>>> http://www.open-mpi.org/mailman/listinfo.cgi/users >>> >> >> >> >> _______________________________________________ >> users mailing list >> >> us...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/users > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users