June 2011 - mpiwg-ft - mailman3.cels.anl.gov

[Mpi3-ft] error vs fault
by Darius Buntinas 23 Jun '11

23 Jun '11

I feel it would be easier to explain/understand FT related things if we distinguished between errors and faults, i.e., an error is a detected fault. I'd like to take a stab at running through the whole standard to make these changes. But before I spend time on this I'd like to make sure people are OK with it. I believe I'd have to be done with this by Wednesday. (right Josh?) Does anyone have objections? -d

1 0

Re: [Mpi3-ft] NUM_STATE vs STATE_SIZE
by Josh Hursey 23 Jun '11

23 Jun '11

My only concern is that "MPI_*_VALIDATE_STATE_SIZE" makes me think about the 'size of the state object' instead of 'the number of states being returned'. But I'm fine with the change if others are. I think there was also a more general concern about the validate function names - making them read a bit better. But I cannot find a clear note about this issue in particular. -- Josh On Thu, Jun 23, 2011 at 2:46 PM, Darius Buntinas <buntinas(a)mcs.anl.gov> wrote: > > A few calls ago, we discussed whether to use MPI_*_VALIDATE_NUM_STATE() vs MPI_*_VALIDATE_STATE_SIZE() to return the number of entries that would be returned by MPI_*_VALIDATE_GET_STATE(). > > Doing a string search through the standard reveals that there are no matches for "_num" outside other than in the FT chapter for MPI_*_VALIDATE_NUM_STATE. ï¿½On the other hand, there are many matches for "_SIZE", e.g., MPI_COMM_SIZE, MPI_PACK_SIZE, MPI_TYPE_SIZE and MPI_GROUP_SIZE. > > So based on precedence (and because I like it better), I think we should use MPI_*_VALIDATE_STATE_SIZE. > > Comments? > > -d > > > _______________________________________________ > mpi3-ft mailing list > mpi3-ft(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft > > -- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey

1 0

[Mpi3-ft] NUM_STATE vs STATE_SIZE
by Darius Buntinas 23 Jun '11

23 Jun '11

A few calls ago, we discussed whether to use MPI_*_VALIDATE_NUM_STATE() vs MPI_*_VALIDATE_STATE_SIZE() to return the number of entries that would be returned by MPI_*_VALIDATE_GET_STATE(). Doing a string search through the standard reveals that there are no matches for "_num" outside other than in the FT chapter for MPI_*_VALIDATE_NUM_STATE. On the other hand, there are many matches for "_SIZE", e.g., MPI_COMM_SIZE, MPI_PACK_SIZE, MPI_TYPE_SIZE and MPI_GROUP_SIZE. So based on precedence (and because I like it better), I think we should use MPI_*_VALIDATE_STATE_SIZE. Comments? -d

1 0

[Mpi3-ft] RTS - 4th Draft
by Josh Hursey 22 Jun '11

22 Jun '11

I just uploaded the 4th draft version of the RTS proposal to the wiki that reflects the notes from todays teleconf. Change log at bottom. https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/ft/run_through_stabilizat… On the teleconf we decided to meet again on Monday, June 27 at 11:30 am EDT - in addition to meeting on Wednesday, June 29 at the regular time. The document should be ready to ship by the Wednesday meeting. We stopped at section 17.9 on todays teleconf. A couple of issues came up on the teleconf that other might want to comment on: - In Section 17.6.2, we need to cleanup the wording here a bit to make it clear about the possible impact on implementations. In particular, does this require that all MPI operations have a check for initialized in them? and how could that check be used to provide the necessary semantics. - Folks were wondering about what we should say regarding how long MPI_Comm_kill should block considering it is a non-collective operation. I added a sentence to the document, not sure if it is good or not but it is a place to start. Anyone have other suggestions? Thanks, Josh Change Log * 17.5.2: Add forward reference to collective chapter in MPI_Comm_validate_all explanation of collective active. * 17.6: Minor wording cleanup * 17.6.1: Tense correction on 'error classes' * 17.6.2: Loosen up the wording a bit 'should try' and add an advice to implementors regarding critical errors. * 17.6.3: Clarify that this is a local operation (non-collective). Also clarify 'process group' * 17.6.3: If the MPI_Comm_kill operation cannot provide the necessary semantics then it should return MPI_ERR_UNSUPPORTED_OPERATION. * 17.6.3: Add a first try at some language about how long MPI_Comm_kill will block - must block until locally finished, may block until globally finished. * 17.7.1: 'effect' -> 'affect' * 17.7.1: Slight clarification improvement to the use of MPI_Comm_validate_set_state_null on an inactive persistent request. * 17.7.1: 'effect' -> 'affect' * 17.8: Clarify advice to implementors * 17.8: Highlight 'undefined' for buffers. * 17.8.1: 'therefor' -> 'therefore' * 17.8.2: Re-order the sentence so it reads a bit better * 17.8.3: Add back reference to Validation operations regarding overlapping collectives with validate_all. * 17.8.4: Bcast Example - Add citation for 'recovery blocks' * 17.9.2: Add missing word ("group") to 'Rationale' * 17.9.2: Clarify that MPI_Comm_is_collectively_active is a local operation. * 17.12.1: Add back reference to window validation operations to advice to users. * 17.13: Put the question in an 'e.g.,' parenthetical. * 17.13.1: Add a bit more clarification to Rationale regarding semantics for MPI_File_close. -- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey

1 0

[Mpi3-ft] FT WG Teleconf Today @ Noon EDT
by Josh Hursey 22 Jun '11

22 Jun '11

Reminder that we have a teleconf today, Wed 6/22/2011, at 12 noon EDT. Date: June 22, 2011 Time: 12 noon EDT/New York Dial-in information: 877-801-8130 Code: 1044056 Agenda: - Run-through stabilization proposal review - https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/ft/run_through_stabilizat… - https://svn.mpi-forum.org/trac/mpi-forum-web/attachment/wiki/ft/run_through… Next Meetings: * June 22, 2011 (Teleconf) * June 29, 2011 (Teleconf) * July 6, 2011 (Teleconf) * July 18, 2011 (MPI Forum Meeting) * Aug. 3, 2011 (Teleconf) * Aug. 17, 2011 (Teleconf) * Aug. 31, 2011 (Teleconf) * Sept. 14, 2011 (Teleconf) * Sept. 22, 2011 (MPI Forum Meeting) Respond to the list if you have any additional topics that you would like to discuss. -- Josh -- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey

1 0

[Mpi3-ft] RTS Proposal - 3rd Draft
by Josh Hursey 20 Jun '11

20 Jun '11

I updated the draft in SVN and on the Wiki page from today's teleconf. Change log at bottom. We changed the name of the global file handle validation functions from MPI_File_validate_sync to MPI_File_validate_all to match the naming convention for groups/communicators/windows. However, the 'Sync' is a reminder that the operation not only agrees upon a global list of failed processes, but also synchronizes the file to disk. How do folks feel about this change? Should we change it back? -- Josh Files can be found here: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/ft/run_through_stabilizat… https://svn.mpi-forum.org/trac/mpi-forum-web/attachment/wiki/ft/run_through… Change Log: * 17.5: Added italics to 'local' and 'global' list upon first introduction. * 17.5: Renamed the 'type' parameter to a more descriptive 'mask' * 17.5.1: Changed wording for MPI_Group_validate to that which was posted by Darius on the list * 17.5.1: Add a short description of the 'mask' in MPI_Group_validate_get_num_state * 17.5.2: Update second sentence of MPI_Comm_validate_all paragraph to be a bit more clear. * 17.5.2: Update fourth sentence of MPI_Comm_validate_all paragraph to be a bit more clear. * 17.5.2: Added a couple sentences about outstanding collective across a MPI_Comm_validate_all * 17.5.4: Changed "MPI_File_validate_sync" to "MPI_File_validate_all" to be consistent with the naming convention * 17.5.4: Changed 'Advice to implementors' to 'Advice to users' for the 'sync-barrier-sync' advice. -- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey

1 0

Re: [Mpi3-ft] Text tweaks
by Darius Buntinas 20 Jun '11

20 Jun '11

New wording based on comments during the call. -d MPI_GROUP_VALIDATE is a process local operation that takes a snapshot of the local list of known failed processes for the specified process group. If any new failures were added to the local list (either since the group was created or the last call to MPI_GROUP_VALIDATE), then newfailures will be set to true, otherwise it will be set to false. The following accessor functions are used to query the snapshot of the local known failed processes taken with MPI_GROUP_VALIDATE. Note the snapshot only change when MPI_GROUP_VALIDATE is called. This means that queries of the state of a process using the following accessor functions will always return the same value until MPI_GROUP_VALIDATE is called again. Rationale. The constant state requirement between calls MPI_GROUP_VALIDATE and similar update validation operations allows the application to manage thread safety while querying the snapshot of local list of processes (maintained by the MPI implementation). (End of rationale.)

1 0

Re: [Mpi3-ft] FT WG Teleconf Today @ Noon EDT
by Josh Hursey 20 Jun '11

20 Jun '11

Yes. Sorry I forgot to send out the reminder this morning. Call-in info below: Time: 11:30 am EDT/New York Dial-in information: 877-801-8130 Code: 1044056 -- Josh On Mon, Jun 20, 2011 at 11:16 AM, Howard Pritchard <howardp(a)cray.com> wrote: > Hi Josh, > > Are you still planning to have this telecon? > > Howard > > Josh Hursey wrote: >> On the call today we decided to meet again on Monday, June 20 at 11:30 >> am EDT/New York for an hour to continue working through the RTS >> proposal. I'll send out a reminder the morning of. We will also meet >> on Wednesday, June 22 at the normal time to continue the discussion. >> >> Thanks, >> Josh >> >> On Wed, Jun 15, 2011 at 9:00 AM, Josh Hursey <jjhursey(a)open-mpi.org> wrote: >>> Reminder that we have a teleconf today, Wed 6/15/2011, at 12 noon EDT. >>> >>> Date: June 15, 2011 >>> Time: 12 noon EDT/New York >>> Dial-in information: >>> 877-801-8130 >>> Code: 1044056 >>> >>> Agenda: >>> - Run-through stabilization proposal review >>> ï¿½- https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/ft/run_through_stabilizat… >>> ï¿½- https://svn.mpi-forum.org/trac/mpi-forum-web/attachment/wiki/ft/run_through… >>> >>> Next Meetings: >>> * June ï¿½15, 2011 (Teleconf) >>> * June ï¿½22, 2011 (Teleconf) >>> * June ï¿½29, 2011 (Teleconf) >>> * July ï¿½ 6, 2011 (Teleconf) >>> * July ï¿½18, 2011 (MPI Forum Meeting) >>> * Aug. ï¿½ 3, 2011 (Teleconf) >>> * Aug. ï¿½17, 2011 (Teleconf) >>> * Aug. ï¿½31, 2011 (Teleconf) >>> * Sept. 14, 2011 (Teleconf) >>> * Sept. 22, 2011 (MPI Forum Meeting) >>> >>> Respond to the list if you have any additional topics that you would >>> like to discuss. >>> >>> -- Josh >>> >>> >>> -- >>> Joshua Hursey >>> Postdoctoral Research Associate >>> Oak Ridge National Laboratory >>> http://users.nccs.gov/~jjhursey >>> >> >> >> > > > -- > Howard Pritchard > Software Engineering > Cray, Inc. > _______________________________________________ > mpi3-ft mailing list > mpi3-ft(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft > > -- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey

1 0

[Mpi3-ft] Text tweaks
by Darius Buntinas 20 Jun '11

20 Jun '11

Hi Josh and group, I took a stab at rewording the MPI_GROUP_VALIDATE description (starting at line 31 on standard page 541). I think this makes the distinction between _VALIDATE function and _VALIDATE_GET_ functions clearer. If we like this wording we'll probably need to update the descriptions of the other _VALIDATE{_ALL} functions. Let me know what you think. -d MPI_GROUP_VALIDATE is a process local operation that takes a snapshot of the local list of known failed processes for the specified process group. If any new failures were added to the local list (either since the group was created or the last call to MPI_GROUP_VALIDATE), then newfailures will be set to true, otherwise it will be set to false. The following accessor functions are used to query the snapshot of the local known failed processes taken with MPI_GROUP_VALIDATE. Note the snapshot will not change between calls to MPI_GROUP_VALIDATE. This means that queries of the state of a process using the following accessor functions will always return the same value until MPI_GROUP_VALIDATE is called again. Rationale. The constant state requirement between calls MPI_GROUP_VALIDATE and similar update validation operations allows the application to manage thread safety while querying the snapshot of local list of processes (maintained by the MPI implementation). (End of rationale.)

1 0

Re: [Mpi3-ft] FT WG Teleconf Today @ Noon EDT
by Howard Pritchard 20 Jun '11

20 Jun '11

Hi Josh, Are you still planning to have this telecon? Howard Josh Hursey wrote: > On the call today we decided to meet again on Monday, June 20 at 11:30 > am EDT/New York for an hour to continue working through the RTS > proposal. I'll send out a reminder the morning of. We will also meet > on Wednesday, June 22 at the normal time to continue the discussion. > > Thanks, > Josh > > On Wed, Jun 15, 2011 at 9:00 AM, Josh Hursey <jjhursey(a)open-mpi.org> wrote: >> Reminder that we have a teleconf today, Wed 6/15/2011, at 12 noon EDT. >> >> Date: June 15, 2011 >> Time: 12 noon EDT/New York >> Dial-in information: >> 877-801-8130 >> Code: 1044056 >> >> Agenda: >> - Run-through stabilization proposal review >> - https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/ft/run_through_stabilizat… >> - https://svn.mpi-forum.org/trac/mpi-forum-web/attachment/wiki/ft/run_through… >> >> Next Meetings: >> * June 15, 2011 (Teleconf) >> * June 22, 2011 (Teleconf) >> * June 29, 2011 (Teleconf) >> * July 6, 2011 (Teleconf) >> * July 18, 2011 (MPI Forum Meeting) >> * Aug. 3, 2011 (Teleconf) >> * Aug. 17, 2011 (Teleconf) >> * Aug. 31, 2011 (Teleconf) >> * Sept. 14, 2011 (Teleconf) >> * Sept. 22, 2011 (MPI Forum Meeting) >> >> Respond to the list if you have any additional topics that you would >> like to discuss. >> >> -- Josh >> >> >> -- >> Joshua Hursey >> Postdoctoral Research Associate >> Oak Ridge National Laboratory >> http://users.nccs.gov/~jjhursey >> > > > -- Howard Pritchard Software Engineering Cray, Inc.

1 0