Per your mail on the devel list, we'll follow up there. Many thanks!
On Apr 29, 2009, at 1:09 PM, Brian Blank wrote:
To Whom This May Concern:
I've been trying to dig a little deeper into this problem and found
some additional information.
First, the stack trace for the ABR and ABW were different. The ABR
problem occurred in datatype_pack.h while the ABW problem occurred
in datatype_unpack.h. The problem appears to be the same still.
Both errors are occurring during a call to MEMCPY_CSUM().
I also found there are two different variables playing into this
bug. There is _copy_blength and _copy_count. At the top of the
function, both of these variables are set to 2 bytes for MPI_SHORT,
4 bytes for MPI_LONG, and 8 bytes for MPI_DOUBLE. Then, these
variables are multiplied together to get the size of the memcpy().
Unfortunetly, the correct size are either of these variables before
they are squared. There sometimes appears to be an optimization
where if two variables are next to each other, they are sometimes
converted into a MPI_BYTE where the size is also incorrectly taking
these squared values into consideration.
I wrote a small test program to illustrate the problem and attached
it to this email. First, I configured openmpi 1.3.2 with the
following options:
./configure --prefix=/myworkingdirectory/openmpi-1.3.2.local --
disable-mpi-f77 --disable-mpi-f90 --enable-debug --enable-mem-debug
--enable-mem-profile
I then modified datatype_pack.h & datatype_unpack.h located in
openmpi-1.3.2/ompi/datatype directory in order to produce additional
debugging output. The new versions are attached to this email.
Then, I executed "make all install"
Then, I write the attached test.c program. You can find the output
below. The problems appear in red.
0: sizes '3' '4' '8' '2'
0: offsets '0' '4' '8' '16'
0: addresses '134510640' '134510644' '134510648' '134510656'
0: name='MPI_CHAR' _copy_blength='3' orig_copy_blength='1'
_copy_count='3' _source='134510640'
0: name='MPI_LONG' _copy_blength='16' orig_copy_blength='4'
_copy_count='4' _source='134510644'
0: name='MPI_DOUBLE' _copy_blength='64' orig_copy_blength='8'
_copy_count='8' _source='134510648'
0: name='MPI_SHORT' _copy_blength='4' orig_copy_blength='2'
_copy_count='2' _source='134510656'
0: one='22' two='222' three='33.300000' four='44'
1: sizes '3' '4' '8' '2'
1: offsets '0' '4' '8' '16'
1: addresses '134510640' '134510644' '134510648' '134510656'
1: name='MPI_CHAR' _copy_blength='3' orig_copy_blength='1'
_copy_count='3' _destination='134510640'
1: name='MPI_LONG' _copy_blength='16' orig_copy_blength='4'
_copy_count='4' _destination='134510644'
1: name='MPI_DOUBLE' _copy_blength='64' orig_copy_blength='8'
_copy_count='8' _destination='134510648'
1: name='MPI_SHORT' _copy_blength='4' orig_copy_blength='2'
_copy_count='2' _destination='134510656'
1: one='22' two='222' three='33.300000' four='44'
You can see from the output that the MPI_Send & MPI_Recv functions
are reading or writing too much data from my structure, causing an
overflow condition to occur. I believe this is causing my
application to crash.
Any help on this problem would be appreciated. If there is anything
else that you need from me, just let me know.
Thanks,
Brian
On Tue, Apr 28, 2009 at 9:58 PM, Brian Blank <brianbl...@gmail.com>
wrote:
To Whom This May Concern:
I am having problems with an OpenMPI application I am writing on the
Solaris/Intel AMD64 platform. I am using OpenMPI 1.3.2 which I
compiled myself using the Solaris C/C++ compiler.
My application was crashing (signal 11) inside a call to malloc()
which my code was running. I thought it might be a memory overflow
error that was causing this, so I fired up Purify. Purify found
several problems inside the the OpenMPI library. I think one of the
errors is serious and might be causing the problems I was looking for.
The serious error is an Array Bounds Write (ABW) which occurred 824
times through 312 calls to MPI_Recv(). This error means that
something inside the OpenMPI library is writing to memory where it
shouldn't be writing to. Here is the stack trace at the time of
this error:
Stack Trace 1 (Occurred 596 times)
memcpy rtlib.o
unpack_predefined_data [datatype_unpack.h:41]
MEMCPY_CSUM( _destination, *(SOURCE), _copy_blength, (CONVERTOR) );
ompi_generic_simple_unpack [datatype_unpack.c:419]
ompi_convertor_unpack [convertor.c:314]
mca_pml_ob1_recv_frag_callback_match [pml_ob1_recvfrag.c:218]
mca_btl_sm_component_progress [btl_sm_component.c:427]
opal_progress [opal_progress.c:207]
opal_condition_wait [condition.h:99]
<Writing 64 bytes to 0x821f738 in heap (16 bytes at 0x821f768
illegal).>
<Address 0x821f738 is 616 bytes into a malloc'd block at 0x821f4d0
of 664 bytes.>
Stack Trace 2 (Occurred 228 times)
memcpy rtlib.o
unpack_predefined_data [datatype_unpack.h:41]
MEMCPY_CSUM( _destination, *(SOURCE), _copy_blength, (CONVERTOR) );
ompi_generic_simple_unpack [datatype_unpack.c:419]
ompi_convertor_unpack [convertor.c:314]
mca_pml_ob1_recv_request_progress_match [pml_ob1_recvreq.c:624]
mca_pml_ob1_Recv_req_start [pml_ob1_recvreq.c:1008]
mca_pml_ob1_recv [pml_ob1_irecv.c:103]
MPI_Recv [precv.c:75]
<Writing 64 bytes to 0x821f738 in heap (16 bytes at 0x821f768
illegal).>
<Address 0x821f738 is 616 bytes into a malloc'd block at 0x821f4d0
of 664 bytes.>
I'm not that familiar with the under workings of the openmpi
library, but I tried to debug it anyway. I noticed that it was
copying a lot of extra bytes for MPI_LONG and MPI_DOUBLE types. On
my system, MPI_LONG should have been 4 bytes, but was copying 16
bytes. Also, MPI_DOUBLE should have been 8 bytes, but was copying
64 bytes. It seems the _copy_blength variable was being set to
high, but I'm not sure why. The above error also shows 64 bytes
being read, where my debugging shows a 64 byte copy for all
MPI_DOUBLE types, which I feel should have been 8 bytes. Therefore,
I really believe _copy_blength is being set to high.
Interestingly enough, the call to MPI_Isend() was generating an ABR
(Array Bounds Read) error in the exact same line of code. The ABR
error sometimes can be fatal if the bytes being read is not a legal
address, but the ABW error is usually a much more fatal error
because it is definitely writing into memory that is probably used
for something else. I'm sure that if we fix the ABW error, the ABR
error should fix itself too as it's the same line of code.
Purify also found 14 UMR (Uninitialized memory read) errors inside
the OpenMPI library. Sometimes this can be really bad if there are
any decisions being made as a result of reading this memory
location. But for now, let's solve the serious error I reported
above first, then I will send you the UMR errors next.
Any help you can provide would be greatly appreciated.
Thanks,
Brian
<
datatype_pack
.h
>
<
datatype_unpack
.h><test.c>_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users
--
Jeff Squyres
Cisco Systems
