Charm++ parallel programming system

[Gengbin Zheng <zhenggb AT gmail.com>]

This could have something to do with the recent change to the default array mapping. Only yesterday I noticed this bug for checkpoint/restart, and I asked Abhinav to fix it.

Phil, is that in the release candidate?

Gengbin

On Tue, Aug 31, 2010 at 6:00 PM, Tom Quinn <trq AT astro.washington.edu> wrote:

I'm still having problems with restarts from checkpoints: The symptom
seems to be that the restart entry executes fine until it calls a proxy
broadcast: only 4055 out of 4096 elements get the broadcast, then ChaNGa
dies soon afterwards.  The restart has 5 empty "arr_*.dat" files.  This is
running the net-linux-x86_64-ibverbs-icc on 512 cores of the NASA Pleiades
machine.

Tom Quinn       Astronomy, University of Washington
Internet:       trq AT astro.washington.edu
Phone:          206-685-9009

On Mon, 30 Aug 2010, Phil Miller wrote:

> I've just incorporated an additional bugfix applicable to checkpoint
> restart. It's in the repository tagged as charm-6.2.2-pre3.
>
> On Mon, Aug 30, 2010 at 12:52, Phil Miller <mille121 AT illinois.edu> wrote:
>> A release candidate of Charm 6.2.2 is available from the Git
>> repository (git checkout charm-6.2.2-pre2), with the following changes
>> from 6.2.1:
>>
>> FEM: prepend string to timestamp files for compatibility with newer paraviews
>> Chare Array default mapping: fix bugs on late insertion and apparent imbalance
>> Makefile: Avoid accidental wrong generation of pup_f.f90
>> Chkpt: account for processors with 0 objects
>> add a test for restarting on a smaller number of pes.
>> fixed default array map in case when check/restart happens on
>> different processors. The binsize has to be updated.
>> icc: Staticly link Intel's libraries on all versions >=9
>> CkMulticast Reductions: Set result message reference number to userFlag
>> Reductions: Set the reference number on result messages to the userFlag
>> NetFEM: add prefix to timestep filenames for Paraview 3.x
>> Fall back on gethostname if a node has multiple IP addresses
>> bluegenep: Update compilers
>> bluegenep: fix path for XLF
>> CPU affinity on mpi-crayxt-smp: correct getXTNodeID calls for SMP
>> mpi-crayxt-smp: fix cputopology to account for multiple cores per node
>> xlC: update linker flags for blueprint
>> xlC 64-bit: link with big library TOC
>> trace-summary: Bugfixes
>> Change timestep filename format to work with paraview 3.x
>> CPU Affinity +pemap: fix a buffer overfow bug.
>> Docs: note slowness in CPU topology gathering from DNS issues
>> xlc: don't pass archaic -qstaticinline
>> Socket Routines: use getifaddrs (when available) for getting local IP.
>> configure: check if cp -p works
>> LAPI: Only copy argv in PEs that are not rank 0 in their process
>> configure: Drop archaic test on C library version
>> Fix the reading of processors lists after +pemap block addition
>> Increment version number to 6.2.2
>>
>>
>> Collectively, these improve support for various architectures, fix
>> bugs in some newer features, and offer some minor improvements for
>> testing and development work.
>>
>> Please test this code and post any discovered problems to the list.
>> Barring any new issues, it will be released as 6.2.2 in the next day
>> or two.
>>
>
_______________________________________________
charm mailing list
charm AT cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/charm