December 2008 - Mpich2-core - mailman3.cels.anl.gov

char* + offset_t
by Darius Buntinas 18 Dec '08

18 Dec '08

Intel found a bug where we were adding a DLOOP_Offset to a char*, and we had cast the DLOOP_Offset to an int, leading to a truncation on a 64-bit machine. e.g., in src/mpid/common/datatype/mpid_segment.c:369 else if (last_idx >= 0 && (last_end == ((char *) bufp + (int) rel_off))) I changed the int cast to a cast to MPI_Pint, which should address their problem and also any compiler warnings you would get due to adding a 64-bit DLOOP_Offset to a 32-bit char* (like on bgp). My question is whether we need an overflow test similar to MPID_Ensure_Aint_fits_in_pointer(). The ensure_aint test is done in the routine in mpid_segment.c, but I think it's the wrong test (it didn't catch Intel's bug). It seems to me that we want MPID_Ensure_DLOOP_Offset_fits_in_pointer(). Does this make sense? Thanks, -d

4 16

Re: [mpich2-core] [mpi-all-commits] r3663 -mpich2/branches/dev/NemesisOnWindows/test/mpi/coll
by Jayesh Krishna 09 Dec '08

09 Dec '08

Spoke with Darius again... It looks like the original test code had a bug, so the right fix should be - ((int *)cinPtr) += matSize*matSize; - ((int *)coutPtr) += matSize*matSize; + cinPtr = (int *)cinPtr + matSize*matSize; + coutPtr = (int *)coutPtr + matSize*matSize; Will fix this in the trunk after testing. I have to figure out how I can use multiple arrays in the test/red4 code... -Jayesh -----Original Message----- From: mpi-all-commits-bounces(a)mcs.anl.gov [mailto:[email protected]] On Behalf Of jayesh(a)mcs.anl.gov Sent: Tuesday, December 09, 2008 1:47 PM To: mpi-all-commits(a)mcs.anl.gov Subject: [mpi-all-commits] r3663 -mpich2/branches/dev/NemesisOnWindows/test/mpi/coll Author: jayesh Date: 2008-12-09 13:46:42 -0600 (Tue, 09 Dec 2008) New Revision: 3663 Modified: mpich2/branches/dev/NemesisOnWindows/test/mpi/coll/red4.c Log: Fixing bug in r3661. The cast had to be (char *). Also some compilers might complain with typecasting and incrementing an lvalue in += . Review@Darius Modified: mpich2/branches/dev/NemesisOnWindows/test/mpi/coll/red4.c =================================================================== --- mpich2/branches/dev/NemesisOnWindows/test/mpi/coll/red4.c 2008-12-09 18:09:31 UTC (rev 3662) +++ mpich2/branches/dev/NemesisOnWindows/test/mpi/coll/red4.c 2008-12-09 19:46:42 UTC (rev 3663) @@ -51,8 +51,8 @@ cout[j+i*matSize] = tempCol[i]; } } - ((int *)cinPtr) += matSize*matSize; - ((int *)coutPtr) += matSize*matSize; + cinPtr = (char *)cinPtr + matSize*matSize; + coutPtr = (char *)coutPtr + matSize*matSize; } }

1 0

Re: [mpich2-core] Fwd: http://s08.idav.ucdavis.edu/munshi-opencl.pdf
by Anthony Chan 08 Dec '08

08 Dec '08

How about NPB-MZ benchmark ? We can vary the OpenMP thread vs process to see how performance changes. Darius, is nemesis supported on this machine ? A.Chan ----- "Rusty Lusk" <lusk(a)mcs.anl.gov> wrote: > Cool. What do we have in the way of hybrid tests for performance? > > On Friday,Dec 5, 2008, at 5:51 PM, Darius Buntinas wrote: > > > > > I've installed 1.0.8 in /home/buntinas/sun4/mpich2-1.0.8/ > > > > It passed all but two threaded tests that I haven't looked into > yet. > > I'm still working on getting 1.1a compiled. > > > > -d > > > > On 12/05/2008 11:30 AM, Rusty Lusk wrote: > >> Someone should try setting up MPICH2 on the new 64-core sun. Ralph > > >> has > >> already done some successful shared-memory-with-processes > >> experiments, > >> including cpi, but without MPI. > >> > >> Begin forwarded message: > >> > >>> *From: *Narayan Desai <desai(a)mcs.anl.gov > <mailto:[email protected]>> > >>> *Date: *December 4, 2008 4:47:30 PM CST > >>> *To: *Rusty Lusk <lusk(a)mcs.anl.gov <mailto:[email protected]>> > >>> *Cc: *Ralph Butler <rbutler(a)mtsu.edu <mailto:[email protected]>> > >>> *Subject: **Re: http://s08.idav.ucdavis.edu/munshi-opencl.pdf* > >>> > >>> On Thu, 4 Dec 2008 16:02:56 -0600 Rusty Lusk wrote: > >>> > >>> Rusty> They are here. I don't think Narayan has turned them > >>> loose yet, > >>> Rusty> though. > >>> > >>> ralph, they aren't entirely ready for prime time yet, but one is > up > >>> enough to play with (and build mpich2 on). try logging into > >>> niagara1.mcs.anl.gov from bblogin, and give things a spin. > >>> -nld > >>

2 1

Fwd: http://s08.idav.ucdavis.edu/munshi-opencl.pdf
by Rusty Lusk 07 Dec '08

07 Dec '08

Someone should try setting up MPICH2 on the new 64-core sun. Ralph has already done some successful shared-memory-with-processes experiments, including cpi, but without MPI. Begin forwarded message: > From: Narayan Desai <desai(a)mcs.anl.gov> > Date: December 4, 2008 4:47:30 PM CST > To: Rusty Lusk <lusk(a)mcs.anl.gov> > Cc: Ralph Butler <rbutler(a)mtsu.edu> > Subject: Re: http://s08.idav.ucdavis.edu/munshi-opencl.pdf > > On Thu, 4 Dec 2008 16:02:56 -0600 Rusty Lusk wrote: > > Rusty> They are here. I don't think Narayan has turned them loose > yet, > Rusty> though. > > ralph, they aren't entirely ready for prime time yet, but one is up > enough to play with (and build mpich2 on). try logging into > niagara1.mcs.anl.gov from bblogin, and give things a spin. > -nld

3 5

FW: LANS Informal Seminar Talks Network
by Rajeev Thakur 02 Dec '08

02 Dec '08

-----Original Message----- From: seminars-bounces(a)lists.mcs.anl.gov [mailto:[email protected]] On Behalf Of Sven Leyffer Sent: Tuesday, December 02, 2008 6:12 PM To: Bob Eisenberg; Dimitrov, Stanko B.; Kenneth L. Judd; seminars(a)mcs.anl.gov Subject: LANS Informal Seminar Talks Network More ferocious than Alien vs. Predator; more moving than Kramer vs. Kramer; and funnier than When Harry Met Sally ... Only at Argonne: AD meets MPI! ======================================================================= Who: Jean Utke What: Adjoinable MPI When: Wednesday at 3pm Where: MCS Room A-261 Goodies: Boyana Coffee: Caribou's Christmas Special ======================================================================= Upcoming Seminars: December 10, "Topics in Economics and Complementarity" Todd Munson December 17, "Nonlinear Least Squares" Stefan Wild Web-page: http://www.mcs.anl.gov/research/LANS/events/listn/

1 0