Known issues with the MPICH2-1.0.6-rc
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've compiled a list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything! Thanks. -- Pavan -- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji MPICH2-1.0.6-RC1 ---------------- 1. Abort test fails with Sock/MPD. 2. Gforker drops parts of stdout once in a while. 3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels. 4. SPARC SHM tests with more processes than processors (2 on atlantis) fail. This is likely because the OS is not able to schedule processes too well during over-subscription. 5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs. 6. Threaded alltoall fails in SMPD.
I missed the manyconnect test for SMPD. I've added it in and committed the file into mpich2tasks. -- Pavan On 08/15/2007 03:45 PM, Pavan Balaji wrote:
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've compiled a list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything!
Thanks.
-- Pavan
------------------------------------------------------------------------
MPICH2-1.0.6-RC1 ----------------
1. Abort test fails with Sock/MPD.
2. Gforker drops parts of stdout once in a while.
3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels.
4. SPARC SHM tests with more processes than processors (2 on atlantis) fail. This is likely because the OS is not able to schedule processes too well during over-subscription.
5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs.
6. Threaded alltoall fails in SMPD.
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
Now threaded alltoall works with SMPD but ctxdup and some others don't. Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 3:48 PM To: [email protected] Subject: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
I missed the manyconnect test for SMPD. I've added it in and committed the file into mpich2tasks.
-- Pavan
On 08/15/2007 03:45 PM, Pavan Balaji wrote:
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've
compiled a
list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything!
Thanks.
-- Pavan
-------------------------------------------------------------- ----------
MPICH2-1.0.6-RC1 ----------------
1. Abort test fails with Sock/MPD.
2. Gforker drops parts of stdout once in a while.
3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels.
4. SPARC SHM tests with more processes than processors (2
on atlantis)
fail. This is likely because the OS is not able to schedule processes too well during over-subscription.
5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs.
6. Threaded alltoall fails in SMPD.
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
Was this fix done today? Last nights tests show that threaded alltoall is failing. -- Pavan On 08/15/2007 04:57 PM, Rajeev Thakur wrote:
Now threaded alltoall works with SMPD but ctxdup and some others don't.
Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 3:48 PM To: [email protected] Subject: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
I missed the manyconnect test for SMPD. I've added it in and committed the file into mpich2tasks.
-- Pavan
On 08/15/2007 03:45 PM, Pavan Balaji wrote:
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've compiled a list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything! Thanks.
-- Pavan
-------------------------------------------------------------- ----------
MPICH2-1.0.6-RC1 ----------------
1. Abort test fails with Sock/MPD.
2. Gforker drops parts of stdout once in a while.
3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels.
4. SPARC SHM tests with more processes than processors (2 on atlantis) fail. This is likely because the OS is not able to schedule processes too well during over-subscription.
5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs.
6. Threaded alltoall fails in SMPD. -- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
Yes, that's what Jayesh was saying. Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 4:59 PM To: [email protected] Subject: Re: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
Was this fix done today? Last nights tests show that threaded alltoall is failing.
-- Pavan
On 08/15/2007 04:57 PM, Rajeev Thakur wrote:
Now threaded alltoall works with SMPD but ctxdup and some others don't.
Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 3:48 PM To: [email protected] Subject: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
I missed the manyconnect test for SMPD. I've added it in and committed the file into mpich2tasks.
-- Pavan
On 08/15/2007 03:45 PM, Pavan Balaji wrote:
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've compiled a list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything! Thanks.
-- Pavan
-------------------------------------------------------------- ----------
MPICH2-1.0.6-RC1 ----------------
1. Abort test fails with Sock/MPD.
2. Gforker drops parts of stdout once in a while.
3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels.
4. SPARC SHM tests with more processes than processors (2 on atlantis) fail. This is likely because the OS is not able to schedule processes too well during over-subscription.
5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs.
6. Threaded alltoall fails in SMPD. -- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
Ok, changed it in the known issues file. If the RC goes out tomorrow, should we aim for migrating to SVN on Monday? -- Pavan On 08/15/2007 05:01 PM, Rajeev Thakur wrote:
Yes, that's what Jayesh was saying.
Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 4:59 PM To: [email protected] Subject: Re: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
Was this fix done today? Last nights tests show that threaded alltoall is failing.
-- Pavan
On 08/15/2007 04:57 PM, Rajeev Thakur wrote:
Now threaded alltoall works with SMPD but ctxdup and some others don't. Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 3:48 PM To: [email protected] Subject: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
I missed the manyconnect test for SMPD. I've added it in and committed the file into mpich2tasks.
-- Pavan
On 08/15/2007 03:45 PM, Pavan Balaji wrote:
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've compiled a list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything! Thanks.
-- Pavan
-------------------------------------------------------------- ----------
MPICH2-1.0.6-RC1 ----------------
1. Abort test fails with Sock/MPD.
2. Gforker drops parts of stdout once in a while.
3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels.
4. SPARC SHM tests with more processes than processors (2 on atlantis) fail. This is likely because the OS is not able to schedule processes too well during over-subscription.
5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs.
6. Threaded alltoall fails in SMPD. -- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
I think we should switch to SVN after the real 1.0.6 release so that we don't break anything before that. Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 5:05 PM To: [email protected] Subject: Re: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
Ok, changed it in the known issues file.
If the RC goes out tomorrow, should we aim for migrating to SVN on Monday?
-- Pavan
On 08/15/2007 05:01 PM, Rajeev Thakur wrote:
Yes, that's what Jayesh was saying.
Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 4:59 PM To: [email protected] Subject: Re: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
Was this fix done today? Last nights tests show that threaded alltoall is failing.
-- Pavan
On 08/15/2007 04:57 PM, Rajeev Thakur wrote:
Now threaded alltoall works with SMPD but ctxdup and some others don't. Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 3:48 PM To: [email protected] Subject: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
I missed the manyconnect test for SMPD. I've added it in and committed the file into mpich2tasks.
-- Pavan
On 08/15/2007 03:45 PM, Pavan Balaji wrote:
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've compiled a list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything! Thanks.
-- Pavan
-------------------------------------------------------------- ----------
MPICH2-1.0.6-RC1 ----------------
1. Abort test fails with Sock/MPD.
2. Gforker drops parts of stdout once in a while.
3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels.
4. SPARC SHM tests with more processes than processors (2 on atlantis) fail. This is likely because the OS is not able to schedule processes too well during over-subscription.
5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs.
6. Threaded alltoall fails in SMPD. -- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
Hi, The fix for alltoall is not committed to CVS yet (while testing the fix I found that ctxdup fails :( with the fix). So currently alltoall and manyconnect fail for SMPD... Regards, Jayesh -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Rajeev Thakur Sent: Wednesday, August 15, 2007 5:07 PM To: [email protected] Subject: RE: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc I think we should switch to SVN after the real 1.0.6 release so that we don't break anything before that. Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 5:05 PM To: [email protected] Subject: Re: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
Ok, changed it in the known issues file.
If the RC goes out tomorrow, should we aim for migrating to SVN on Monday?
-- Pavan
On 08/15/2007 05:01 PM, Rajeev Thakur wrote:
Yes, that's what Jayesh was saying.
Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 4:59 PM To: [email protected] Subject: Re: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
Was this fix done today? Last nights tests show that threaded alltoall is failing.
-- Pavan
On 08/15/2007 04:57 PM, Rajeev Thakur wrote:
Now threaded alltoall works with SMPD but ctxdup and some others don't. Rajeev
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Pavan Balaji Sent: Wednesday, August 15, 2007 3:48 PM To: [email protected] Subject: [mpich2-core] Re: Known issues with the MPICH2-1.0.6-rc
I missed the manyconnect test for SMPD. I've added it in and committed the file into mpich2tasks.
-- Pavan
On 08/15/2007 03:45 PM, Pavan Balaji wrote:
It looks like the RC might finally be out tomorrow assuming that we don't see any surprises in the nightly tests tonight. I've compiled a list of issues which we are not dealing with for this RC and need to be fixed eventually. The list is attached with this email. If I've missed something, please let me know -- we don't want to lose track of anything! Thanks.
-- Pavan
-------------------------------------------------------------- ----------
MPICH2-1.0.6-RC1 ----------------
1. Abort test fails with Sock/MPD.
2. Gforker drops parts of stdout once in a while.
3. MPI_Ssend_Overtake tests in the Intel test suite fail on 2.4 kernels due to the poll overhead on these kernels.
4. SPARC SHM tests with more processes than processors (2 on atlantis) fail. This is likely because the OS is not able to schedule processes too well during over-subscription.
5. Datatype performance tests for packing arrays larger than 2 dimensions fail. It looks like this is because the compiler can optimize the application code more easily as compared to the MPICH2 code, where all packing is broken down into 1-D packs.
6. Threaded alltoall fails in SMPD. -- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
-- Pavan Balaji, Mathematics and Computer Science, Argonne National Laboratory Ph: 630.252.3017 http://www.mcs.anl.gov/~balaji
participants (3)
-
Jayesh Krishna -
Pavan Balaji -
Rajeev Thakur