From the meetings website [1], the many tickets appear to be up for votes. See the end of this email for a cut-and-paste list from the agenda. I am currently planning on voting YES for everything except for the Fault Tolerance tickets (#323, #326, #327) and the Helper Threads ticket (#217). Rolf's slides indicate that there may be some other FT tickets up for 1st votes, namely #324. I will vote NO on all of the tickets explicitly enumerated in this paragraph. Are we all on the same page here w.r.t. ANL's position in the Forum? -Dave [1] http://meetings.mpi-forum.org/secretary/2012/05/agenda.php ----8<---- First Votes - For reductions: Grouping MPI_AINT, MPI_OFFSET, MPI_COUNT as Multi-language types ticket #187 C++ binding for Dist_graph_neighbors_count is badly broken ticket #192 Ensure MPI_Dims_create is really suitable for MPI_Cart_create ticket #194 Topology awareness in MPI_Dims_create ticket #195 MPI3 Hybrid Programming: Proposal for Helper Threads ticket #217 MPI_PROC_NULL behavior for MPI_PROBE and MPI_IPROBE not directly defined ticket #256 Functions to query MPI_Info object attached to windows and communicators ticket #271 Add Immediate versions of nonblocking collective I/O routines ticket #273 Update examples to not use deprecated constructs (where possible) ticket #278 Remove C++ bindings ticket #281 MPI_UNWEIGHTED should not be NULL ticket #294 Fix issue with definition of nonblocking in One Sided Chapter ticket #300 Move MPI-2 deprecated functions to new "Removed interfaces" chapter ticket #303 MPI_INIT & MPI_FINALIZE ticket #313 Clarify MPI behavior when multiple MPI processes run in the same address space ticket #310 correct error related to MPI_REQUEST_FREE ticket #317 New FT proposal ticket #323 MPI3 Fault Tolerance - Files ticket #326 MPI3 Fault Tolerance - Dynamic process management ticket #327 fix MPI_PROC_NULL behavior for mprobe/improbe/mrecv/imrecv ticket #328 Second Votes - RMA - Remove sentence on offset in datatypes ticket #308 RMA - Clarify usage of status for request-based RMA operations ticket #309 Allocate a shared memory window ticket #284 Add a new datatype creation routine - MPI_Type_create_hindexed_block ticket #280 Nonblocking Communicator Duplication ticket #168 Remove non-blocking collective C++ bindings ticket #272 Noncollective Communicator Creation ticket #286 Update MPI_Intercomm_create to use collective tag space ticket #305 ----8<----
I think #324 should be OK for us to vote yes, but it's minor so I won't push for it. -d On May 25, 2012, at 5:45 AM, Dave Goodell wrote:
From the meetings website [1], the many tickets appear to be up for votes. See the end of this email for a cut-and-paste list from the agenda.
I am currently planning on voting YES for everything except for the Fault Tolerance tickets (#323, #326, #327) and the Helper Threads ticket (#217). Rolf's slides indicate that there may be some other FT tickets up for 1st votes, namely #324. I will vote NO on all of the tickets explicitly enumerated in this paragraph.
Are we all on the same page here w.r.t. ANL's position in the Forum?
-Dave
[1] http://meetings.mpi-forum.org/secretary/2012/05/agenda.php
----8<---- First Votes - For reductions: Grouping MPI_AINT, MPI_OFFSET, MPI_COUNT as Multi-language types ticket #187 C++ binding for Dist_graph_neighbors_count is badly broken ticket #192 Ensure MPI_Dims_create is really suitable for MPI_Cart_create ticket #194 Topology awareness in MPI_Dims_create ticket #195 MPI3 Hybrid Programming: Proposal for Helper Threads ticket #217 MPI_PROC_NULL behavior for MPI_PROBE and MPI_IPROBE not directly defined ticket #256 Functions to query MPI_Info object attached to windows and communicators ticket #271 Add Immediate versions of nonblocking collective I/O routines ticket #273 Update examples to not use deprecated constructs (where possible) ticket #278 Remove C++ bindings ticket #281 MPI_UNWEIGHTED should not be NULL ticket #294 Fix issue with definition of nonblocking in One Sided Chapter ticket #300 Move MPI-2 deprecated functions to new "Removed interfaces" chapter ticket #303 MPI_INIT & MPI_FINALIZE ticket #313 Clarify MPI behavior when multiple MPI processes run in the same address space ticket #310 correct error related to MPI_REQUEST_FREE ticket #317 New FT proposal ticket #323 MPI3 Fault Tolerance - Files ticket #326 MPI3 Fault Tolerance - Dynamic process management ticket #327 fix MPI_PROC_NULL behavior for mprobe/improbe/mrecv/imrecv ticket #328
Second Votes - RMA - Remove sentence on offset in datatypes ticket #308 RMA - Clarify usage of status for request-based RMA operations ticket #309 Allocate a shared memory window ticket #284 Add a new datatype creation routine - MPI_Type_create_hindexed_block ticket #280 Nonblocking Communicator Duplication ticket #168 Remove non-blocking collective C++ bindings ticket #272 Noncollective Communicator Creation ticket #286 Update MPI_Intercomm_create to use collective tag space ticket #305 ----8<----
OK, I'll switch to voting YES on #324 if it comes up for a vote. -Dave On May 24, 2012, at 5:43 PM CDT, Darius Buntinas wrote:
I think #324 should be OK for us to vote yes, but it's minor so I won't push for it.
-d
On May 25, 2012, at 5:45 AM, Dave Goodell wrote:
From the meetings website [1], the many tickets appear to be up for votes. See the end of this email for a cut-and-paste list from the agenda.
I am currently planning on voting YES for everything except for the Fault Tolerance tickets (#323, #326, #327) and the Helper Threads ticket (#217). Rolf's slides indicate that there may be some other FT tickets up for 1st votes, namely #324. I will vote NO on all of the tickets explicitly enumerated in this paragraph.
Are we all on the same page here w.r.t. ANL's position in the Forum?
-Dave
[1] http://meetings.mpi-forum.org/secretary/2012/05/agenda.php
----8<---- First Votes - For reductions: Grouping MPI_AINT, MPI_OFFSET, MPI_COUNT as Multi-language types ticket #187 C++ binding for Dist_graph_neighbors_count is badly broken ticket #192 Ensure MPI_Dims_create is really suitable for MPI_Cart_create ticket #194 Topology awareness in MPI_Dims_create ticket #195 MPI3 Hybrid Programming: Proposal for Helper Threads ticket #217 MPI_PROC_NULL behavior for MPI_PROBE and MPI_IPROBE not directly defined ticket #256 Functions to query MPI_Info object attached to windows and communicators ticket #271 Add Immediate versions of nonblocking collective I/O routines ticket #273 Update examples to not use deprecated constructs (where possible) ticket #278 Remove C++ bindings ticket #281 MPI_UNWEIGHTED should not be NULL ticket #294 Fix issue with definition of nonblocking in One Sided Chapter ticket #300 Move MPI-2 deprecated functions to new "Removed interfaces" chapter ticket #303 MPI_INIT & MPI_FINALIZE ticket #313 Clarify MPI behavior when multiple MPI processes run in the same address space ticket #310 correct error related to MPI_REQUEST_FREE ticket #317 New FT proposal ticket #323 MPI3 Fault Tolerance - Files ticket #326 MPI3 Fault Tolerance - Dynamic process management ticket #327 fix MPI_PROC_NULL behavior for mprobe/improbe/mrecv/imrecv ticket #328
Second Votes - RMA - Remove sentence on offset in datatypes ticket #308 RMA - Clarify usage of status for request-based RMA operations ticket #309 Allocate a shared memory window ticket #284 Add a new datatype creation routine - MPI_Type_create_hindexed_block ticket #280 Nonblocking Communicator Duplication ticket #168 Remove non-blocking collective C++ bindings ticket #272 Noncollective Communicator Creation ticket #286 Update MPI_Intercomm_create to use collective tag space ticket #305 ----8<----
On Thu, May 24, 2012 at 04:45:17PM -0500, Dave Goodell wrote:
From the meetings website [1], the many tickets appear to be up for votes. See the end of this email for a cut-and-paste list from the agenda.
I am currently planning on voting YES for everything except for the Fault Tolerance tickets (#323, #326, #327) and the Helper Threads ticket (#217). Rolf's slides indicate that there may be some other FT tickets up for 1st votes, namely #324. I will vote NO on all of the tickets explicitly enumerated in this paragraph.
At face value "helper threads" sound like something useful in several I/O contexts. I'm way out of the loop on the discussion though, so I'm sure you have good reasons for voting NO and I don't wish to argue the point. Only curious about the reason. Is the proposal poorly formed? a bad idea in general? poses surprising performance drawbacks? impossible to actually implement?
Are we all on the same page here w.r.t. ANL's position in the Forum?
-Dave
[1] http://meetings.mpi-forum.org/secretary/2012/05/agenda.php
----8<---- First Votes - For reductions: Grouping MPI_AINT, MPI_OFFSET, MPI_COUNT as Multi-language types ticket #187 C++ binding for Dist_graph_neighbors_count is badly broken ticket #192 Ensure MPI_Dims_create is really suitable for MPI_Cart_create ticket #194 Topology awareness in MPI_Dims_create ticket #195 MPI3 Hybrid Programming: Proposal for Helper Threads ticket #217 MPI_PROC_NULL behavior for MPI_PROBE and MPI_IPROBE not directly defined ticket #256 Functions to query MPI_Info object attached to windows and communicators ticket #271 Add Immediate versions of nonblocking collective I/O routines ticket #273 Update examples to not use deprecated constructs (where possible) ticket #278 Remove C++ bindings ticket #281 MPI_UNWEIGHTED should not be NULL ticket #294 Fix issue with definition of nonblocking in One Sided Chapter ticket #300 Move MPI-2 deprecated functions to new "Removed interfaces" chapter ticket #303 MPI_INIT & MPI_FINALIZE ticket #313 Clarify MPI behavior when multiple MPI processes run in the same address space ticket #310 correct error related to MPI_REQUEST_FREE ticket #317 New FT proposal ticket #323 MPI3 Fault Tolerance - Files ticket #326 MPI3 Fault Tolerance - Dynamic process management ticket #327 fix MPI_PROC_NULL behavior for mprobe/improbe/mrecv/imrecv ticket #328
Second Votes - RMA - Remove sentence on offset in datatypes ticket #308 RMA - Clarify usage of status for request-based RMA operations ticket #309 Allocate a shared memory window ticket #284 Add a new datatype creation routine - MPI_Type_create_hindexed_block ticket #280 Nonblocking Communicator Duplication ticket #168 Remove non-blocking collective C++ bindings ticket #272 Noncollective Communicator Creation ticket #286 Update MPI_Intercomm_create to use collective tag space ticket #305 ----8<----
-- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA
On May 25, 2012, at 11:39 PM GMT+09:00, Rob Latham wrote:
On Thu, May 24, 2012 at 04:45:17PM -0500, Dave Goodell wrote:
I am currently planning on voting YES for everything except for the Fault Tolerance tickets (#323, #326, #327) and the Helper Threads ticket (#217).
At face value "helper threads" sound like something useful in several I/O contexts. I'm way out of the loop on the discussion though, so I'm sure you have good reasons for voting NO and I don't wish to argue the point. Only curious about the reason. Is the proposal poorly formed? a bad idea in general? poses surprising performance drawbacks? impossible to actually implement?
That proposal isn't what you think it is. It's supposedly a way for applications to lend threads to the library in order to drive progress on network hardware that needs a lot of CPU in order to drive it at full bandwidth (think BG/P tree network accessed via MMIO). I think you can also use it to accelerate software-based large message allreduce operations (and similar). I suspect that you want it to be background threads for I/O progress when the application is not calling MPI routines, which is not what this proposal is about. Here are a few reasons to vote against it: * poor naming: nobody understands what the hell is going on when you describe it to them with the words used in the proposal * limited utility: it only accelerates a very narrow case, IMO * standardizing total guesses: only one of the several modes of this proposal has concrete optimizations behind it, the rest is just sort of wishing for the future * only one organization really wants it (IBM), which is usually a bad sign -Dave
participants (3)
-
Darius Buntinas -
Dave Goodell -
Rob Latham