Open MPI logo

Hardware Locality Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Hardware Locality Development mailing list

Subject: Re: [hwloc-devel] hwloc on PPC64
From: Jirka Hladky (jhladky_at_[hidden])
Date: 2010-07-11 18:08:40


On Sunday, July 11, 2010 07:57:48 pm Brice Goglin wrote:
> Le 11/07/2010 19:48, Jirka Hladky a écrit :
> > Hi all,
> >
> > I have run into two bugs on PPC64 on 2.6.32 kernel.
> >
> > Version:
> >
> > lt-lstopo 1.0.1
> >
> > BUG #1: No Socket information in lstopo output:
> >
> > ./lstopo
> >
> > Machine (3654MB) + L2 #0 (4096KB)
> >
> > L1 #0 (64KB) + Core #0
> >
> > PU #0 (phys=0)
> >
> > PU #1 (phys=1)
> >
> > L1 #1 (64KB) + Core #1
> >
> > PU #2 (phys=2)
> >
> > PU #3 (phys=3)
> >
> > Fixed in the latest version (tried hwloc-1.1a1r2301.tar.gz)
> > <http://www.open-mpi.org/software/hwloc/nightly/trunk/hwloc-1.1a1r2301.ta
> > r.gz>
>
> In 1.0.1, there's a patch that prevents us from showing invalid socket
> info on old kernels but it also prevents us from showing valid socket
> info on recent kernel. I reverted the commit in trunk (and in the
> upcoming 1.0.2).
Thanks for shading some light into it!

>
> > BUG #2
> >
> > On some PPC64, kernel 2.6.32 I have crash when running
> >
> > $ lstopo a.txt
> >
> > Segmentation fault (core dumped)
> >
> > $ gdb /usr/local/bin/lstopo core.8771
> >
> > Program terminated with signal 11, Segmentation fault.
> >
> > #0 0x00000000100060b4 in .merge ()
> >
> > It appears only on some PPC64 boxes.
> >
> > This issue is also gone with in the latest version (tried
> > hwloc-1.1a1r2301.tar.gz)
> > <http://www.open-mpi.org/software/hwloc/nightly/trunk/hwloc-1.1a1r2301.ta
> > r.gz>
> >
> > I wonder if you are aware of these problems. let me know if you need
> > more details.
>
> If you do "lstopo a.xml" first, does "lstopo --xml a.xml a.txt" crash as
> above? If so, please send a.xml so that I debug this.

$./lstopo --version
lt-lstopo 1.0.1

$./lstopo --xml /tmp/2010-Jul-10_22h14m_results/2.6.32-44.el6.ppc64_OS-
indexing.xml a.txt
Segmentation fault (core dumped)

xml was generated with
lstopo --physical a.xml

Output of command: "lstopo --physical -"
Machine (4096MB)
  NUMANode p#0 (2240MB)
    L1 (64KB) + Core p#0
      PU p#0
      PU p#1
    L1 (64KB) + Core p#2
      PU p#2
      PU p#3
    L1 (64KB) + Core p#4
      PU p#4
      PU p#5
    L1 (64KB) + Core p#6
      PU p#6
      PU p#7
  NUMANode p#1 (1856MB)

Note missing socket.

I will attach:
-xml causing crash (2.6.32-44.el6.ppc64_OS-indexing.xml)
-whole run directory (notice that png, pdf, ... are created (no crash) but are
empty. Others format are OK (check .fig) ) Please notice that hwloc-distrib is
also not working correctly - check CPU_AFFINITY/0008.log for example.
-runtest.sh - script used to create the data.

Let me know if you need more data.

Thanks!
Jirka





  • application/x-shellscript attachment: runtest.sh