[Mpi-forum] MPI-3: MPI_T_ERR_... and MPI_ERR_LASTCODE

Schulz, Martin schulzm at llnl.gov
Tue Jul 24 01:13:54 CDT 2012 

Hi Kathryn, all,

On Jul 23, 2012, at 1:05 PM, Mohror, Kathryn wrote:

> I have some concerns over this.
> 
>> The topic about user-callability of MPI_ERROR_CLASS and MPI_ERROR_STRING
>> outside of the initialized MPI should be revisited in MPI-3.1.
>> This is not evident for the tools' developers because they know
>> the meaning of their MPI_T_ERR... classes and only classes are returned.
> 
> If we are going to treat MPI_T error codes the same way they are in the rest of the standard, does it make sense to say that "only classes are returned" by MPI_T calls?

That's more or less what the last proposal is - we have MPI_T return codes and those should be treated as MPI error classes. I guess we always refrained from calling them error classes or error codes since we put more constraints on them wrt. their influence on the overall execution and because we defined them more closely.

> I think that MPI calls return codes and the codes are translated to classes by the MPI_ERROR_CLASS function.

That's not how it was intended. I agree, though, we didn't formulate this in the most optimal way.

> If MPI_ERROR_CLASS and MPI_ERROR_STRING are not available before MPI_Init, then this may be a hardship on tools. I imagine many tools doing a significant amount of set up before MPI_Init to avoid overhead. However, if there is no information about what the return codes mean, then tools may have difficulty.
> 
> Possibly it makes more sense to treat MPI_T return values differently than MPI calls? Could the MPI_T return values be a set of defined codes instead?

Yes, I agree - that's the intent of the big table at the end of the tools section - fixed return codes for each condition, such that one needs no translation. That's also why I like the currently proposed formulation of having MPI_T return codes that are (only) treated the same as MPI error classes. There were only two reasons why we stuck them at the end into the same number space as MPI error classes or codes: a) it should be possible to return an MPI_T error as a return value of an MPI function (e.g., if the call got intercepted using PMPI and the wrapper uses MPI_T) and b) it was consistent with shifting from MPIT (separate name space) to MPI as the prefix.

Martin



> 
> Does anyone have different insight on error codes vs classes?
> 
> Kathryn
> 
>> -----Original Message-----
>> From: mpi-forum-bounces at lists.mpi-forum.org [mailto:mpi-forum-
>> bounces at lists.mpi-forum.org] On Behalf Of Bronis R. de Supinski
>> Sent: Monday, July 23, 2012 10:37 AM
>> To: Main MPI Forum mailing list
>> Subject: Re: [Mpi-forum] MPI-3: MPI_T_ERR_... and MPI_ERR_LASTCODE
>> 
>> 
>> OK by me.
>> 
>> On Mon, 23 Jul 2012, Dave Goodell wrote:
>> 
>>> I agree, we should make these changes.
>>> 
>>> -Dave
>>> 
>>> On Jul 23, 2012, at 8:55 AM CDT, Schulz, Martin wrote:
>>> 
>>>> Hi Rolf, all,
>>>> 
>>>> I agree with your suggestions and I am in favor of making these changes.
>> Bronis (as well as Kathryn and Dave as the remaining members in the chapter
>> committee for tools) should concur, though.
>>>> 
>>>> Anybody else have any concerns over these changes?
>>>> 
>>>> As for the callability of the error functions, I also agree - let's target those in
>> 3.1. This change is not vital and bit too big for final edits.
>>>> 
>>>> Thanks,
>>>> 
>>>> Martin
>>>> 
>>>> 
>>>> On Jul 23, 2012, at 4:00 AM, Rolf Rabenseifner wrote:
>>>> 
>>>>> Martin et al.,
>>>>> 
>>>>> Consistently with your proposal, I would recommend:
>>>>> 
>>>>> 1) Sect 14.3.9 "Return Codes for the MPI tool information interface"
>>>>> last sentence reads
>>>>> 
>>>>>  All return codes with the prefix MPI_T_ must be unique values and
>>>>>  cannot overlap with any other return values returned by the MPI
>>>>>  implementation.
>>>>> 
>>>>> but should read
>>>>> 
>>>>>> All return codes with the prefix MPI_T_ must be unique values and
>>>>>> cannot overlap with any other
>>>>>  error codes and error classes
>>>>>> returned by the MPI
>>>>>> implementation. Further, they shall be treated as MPI error classes as
>>>>>> defined in Chapter 8.4 and follow the same rules and restrictions,
>>>>>  especially they must satisfy
>>>>> 
>>>>>>>>>> 0 = MPI_SUCCESS < MPI_T_ERR_... \leq MPI_ERR_LASTCODE.
>>>>> 
>>>>> 2) A.1.1 the first three tables are headed by
>>>>> 
>>>>> Error classes
>>>>> Error classes (continued)
>>>>> Return Codes for the MPI tool information interface
>>>>> 
>>>>> and these lines should read
>>>>> 
>>>>> Error classes
>>>>> Error classes (continued)
>>>>> Error classes (continued)
>>>>> 
>>>>> 3) The last line of A.1.1, 2nd table
>>>>> 
>>>>> MPI_ERR_LASTCODE
>>>>> 
>>>>> should move after the last MPI_T_ERR_... code in the 3rd table.
>>>>> 
>>>>> Okay?
>>>>> 
>>>>> The topic about user-callability of MPI_ERROR_CLASS and
>> MPI_ERROR_STRING
>>>>> outside of the initialized MPI should be revisited in MPI-3.1.
>>>>> This is not evident for the tools' developers because they know
>>>>> the meaning of their MPI_T_ERR... classes and only classes are returned.
>>>>> 
>>>>> I would like to have a final "okay" before I execute
>>>>> the changes in chap-appLang.
>>>>> 
>>>>> Best regards
>>>>> Rolf
>>>>> 
>>>>> ----- Original Message -----
>>>>>> From: "Martin Schulz" <schulzm at llnl.gov>
>>>>>> To: "Main MPI Forum mailing list" <mpi-forum at lists.mpi-forum.org>
>>>>>> Sent: Monday, July 23, 2012 1:23:40 AM
>>>>>> Subject: Re: [Mpi-forum] MPI-3: MPI_T_ERR_... and
>> MPI_ERR_LASTCODE
>>>>>> Hi Rolf, all,
>>>>>> 
>>>>>> I thought we had discussed the error semantics of MPI_T in several
>>>>>> meetings/readings and nobody objected, but I generally agree with your
>>>>>> comments below. However, they are not really error classes (at least
>>>>>> in the tools group we/I never thought about them this way), but just
>>>>>> well defined return codes. Nevertheless, they fit the model and
>>>>>> semantics of classes and hence should integrated into the same rules
>>>>>> for consistency. I also agree with Bronis, though, that larger
>>>>>> feedback on this would be good to avoid errors because of rushing it.
>>>>>> 
>>>>>> To keep changes minimal, I would suggest that we only change the
>>>>>> following sentence, which IMHO is sufficient (and would essentially
>>>>>> follow option b):
>>>>>> 
>>>>>> All return codes with the prefix MPI_T_ must be unique values and
>>>>>> cannot overlap with any other return values returned by the MPI
>>>>>> implementation. Further, they shall be treated as MPI error classes as
>>>>>> defined in Chapter 8.4 and follow the same rules and restrictions.
>>>>>> 
>>>>>> This would also make changes in the appendix unnecessary.
>>>>>> 
>>>>>> As for making MPI_ERROR_CLASS and MPI_ERROR_STRING callable
>> before
>>>>>> Init (and then also after finalize), yes that would be very useful,
>>>>>> but is a general issue not only related to MPI_T. If we decide this is
>>>>>> too invasive at this point, I would like to see this at least in 3.1.
>>>>>> 
>>>>>> Martin
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>> On Jul 22, 2012, at 1:46 PM, Rolf Rabenseifner wrote:
>>>>>> 
>>>>>>> About 1. question:
>>>>>>> 
>>>>>>> I changed the headers of the three error code tables in the Annex
>>>>>>> according to Tables 8.1 and 8.2 from "Return codes"
>>>>>>> into "Error classes".
>>>>>>> 
>>>>>>> svn r1505: I used also "Error classes" for the tools table
>>>>>>> svn r1507: I went back to the official ticket 266 text "Return
>>>>>>> codes"
>>>>>>> 
>>>>>>> The Tools group should decide whether
>>>>>>> a. they want to stay with special return codes that are not
>>>>>>> part of the rule in Set. 8.4
>>>>>>> 
>>>>>>> 0 = MPI_SUCCESS < MPI_ERR_... <= MPI_ERR_LASTCODE
>>>>>>> 
>>>>>>> In this case they should change "return code"
>>>>>>> into "error codes are returned" plus noting that
>>>>>>> the routine MPI_ERROR_STRING can be applied
>>>>>>> with the open question, whether MPI_EEROR_STRING
>>>>>>> can be applied before a call to MPI_INIT.
>>>>>>> 
>>>>>>> Similar problem with failed MPI_INIT and analysing its
>>>>>>> returned error code:
>>>>>>> One needs at least MPI_ERROR_CLASS and MPI_ERROR_STRING
>>>>>>> callable before MPI_INIT.
>>>>>>> 
>>>>>>> b. or you may use the same terminology as in Sect. 8.4,
>>>>>>> then new table 14.5 must show "Error classes".
>>>>>>> 
>>>>>>> In this case you should integrate the MPI_T_ERR_...
>>>>>>> into the MPI_ERR_LASTCODE rule.
>>>>>>> 
>>>>>>> You need at least MPI_ERROR_CLASS and MPI_ERROR_STRING
>>>>>>> callable before MPI_INIT.
>>>>>>> 
>>>>>>> I would prefer solution b together combined with the
>>>>>>> minimal solution for failed MPI_INIT, i.e.,
>>>>>>> adding MPI_ERROR_CLASS and MPI_ERROR_STRING to
>>>>>>> the list of routines callable outside of MPI's
>>>>>>> initialization.
>>>>>>> 
>>>>>>> 
>>>>>>> About 2. question: MPI_T_ERR_CANTINIT -_>
>> MPI_T_ERR_CANNOTINIT
>>>>>>> 
>>>>>>> svn r1506: done by Bronis in chap-tools/mpit.tex
>>>>>>> svn r1508: done by Rolf in chap-appLang/appLang-Const.tex
>>>>>>> 
>>>>>>> Rolf
>>>>>>> 
>>>>>>> ----- Original Message -----
>>>>>>>> From: "Bronis R. de Supinski" <bronis at llnl.gov>
>>>>>>>> To: "Main MPI Forum mailing list" <mpi-forum at lists.mpi-forum.org>
>>>>>>>> Cc: "Rolf Rabenseifner" <rabenseifner at hlrs.de>
>>>>>>>> Sent: Sunday, July 22, 2012 10:03:44 PM
>>>>>>>> Subject: Re: [Mpi-forum] MPI-3: MPI_T_ERR_... and
>> MPI_ERR_LASTCODE
>>>>>>>> Rolf:
>>>>>>>> 
>>>>>>>> OK, I am changing MPI_T_ERR_CANTINIT to
>> MPI_T_ERR_CANNOTINIT.
>>>>>>>> 
>>>>>>>> As said before, we need more opinions on the first question.
>>>>>>>> 
>>>>>>>> Bronis
>>>>>>>> 
>>>>>>>> 
>>>>>>>> On Sun, 22 Jul 2012, Bronis R. de Supinski wrote:
>>>>>>>> 
>>>>>>>>> 
>>>>>>>>> Rolf:
>>>>>>>>> 
>>>>>>>>> Re:
>>>>>>>>>> Dear Bronis, Martin, Dave, and Kathryn,
>>>>>>>>>> (Profiling/Tools chapter responsibles)
>>>>>>>>>> 
>>>>>>>>>> (you may forward this to the tools list - I'm not on that list)
>>>>>>>>>> 
>>>>>>>>>> Two important questions:
>>>>>>>>>> 
>>>>>>>>>> 1. Alexander Supalov detected that we did it wrong:
>>>>>>>>>> The MPI_T_ERR_... list must be part of the total error
>>>>>>>>>> list and sorted in before MPI_ERR_LASTCODE.
>>>>>>>>>> 
>>>>>>>>>> I will change this in chap-appLang/appLang-Const.tex
>>>>>>>>>> 
>>>>>>>>>> Please look, if there must be some additional wording on this.
>>>>>>>>>> For example, the last sentence of Section 14.3.9
>>>>>>>>>> 
>>>>>>>>>> "All return codes with the prefix MPI_T_ must be unique values
>>>>>>>>>> and
>>>>>>>>>> cannot overlap with any other return values returned by the MPI
>>>>>>>>>> implementation."
>>>>>>>>> 
>>>>>>>>> The above wording clearly states that the return values
>>>>>>>>> should be consistent with MPI_ERR_* return values. So,
>>>>>>>>> either the above wording needs to change (I do not see
>>>>>>>>> why MPI_T_* return values need to be distinct from
>>>>>>>>> MPI_ERR_* return values since the user should know that
>>>>>>>>> they were using an MPI_T_* function) or we need to adopt
>>>>>>>>> something like what you suggest.
>>>>>>>>> 
>>>>>>>>> I think we need broader opinions to make the decision.
>>>>>>>>> 
>>>>>>>>> I am concerned that wrapping up the document for MPI 3.0
>>>>>>>>> has led to a fast and loose attitude about making broader
>>>>>>>>> changes "in order to get it done." This attitude can
>>>>>>>>> easily lead to suboptimal solutions.
>>>>>>>>> 
>>>>>>>>> Bronis
>>>>>>>>> 
>>>>>>>>> 
>>>>>>>>> 
>>>>>>>>>> may be modified into
>>>>>>>>>> 
>>>>>>>>>> "All return codes with the prefix MPI_T_ must be unique values
>>>>>>>>>> and
>>>>>>>>>> cannot overlap with any other return values returned by the MPI
>>>>>>>>>> implementation and satisfy
>>>>>>>>>> 
>>>>>>>>>> 0 = MPI_SUCCESS < MPI_T_ERR_... <= MPI_ERR_LASTCODE. "
>>>>>>>>>> 
>>>>>>>>>> or
>>>>>>>>>> 0 = MPI_SUCCESS < MPI_ERR_... < MPI_T_ERR_... <=
>>>>>>>>>> MPI_ERR_LASTCODE.
>>>>>>>>>> 
>>>>>>>>>> or
>>>>>>>>>> 0 = MPI_SUCCESS < MPI_ERR_... <= MPI_ERR_LASTCODE <
>>>>>>>>>> MPI_T_ERR_... <= MPI_T_ERR_LASTCODE.
>>>>>>>>>> 
>>>>>>>>>> and the MPI_ERR_LASTCODE and/or MPI_T_ERR_LASTCODE may
>> be also
>>>>>>>>>> repeated
>>>>>>>>>> in Table 14.5 as last entry, as done for the first value
>>>>>>>>>> MPI_SUCCESS.
>>>>>>>>>> 
>>>>>>>>>> This proposal is based on Section 8.4, the sentence
>>>>>>>>>> 
>>>>>>>>>> "The error codes satisfy,
>>>>>>>>>> 0 = MPI_SUCCESS < MPI_ERR_... <= MPI_ERR_LASTCODE: "
>>>>>>>>>> 
>>>>>>>>>> Please confirm, that Alexander is right,
>>>>>>>>>> please include me in your discussion and
>>>>>>>>>> please tell me your result
>>>>>>>>>> that I can do the correct changes in chap-appLang
>>>>>>>>>> and you in chap-tools.
>>>>>>>>>> 
>>>>>>>>>> 2. Qeustion is less important; it is about naming:
>>>>>>>>>> Alexander also noticed that your names do not fit to the names
>>>>>>>>>> used in the past:
>>>>>>>>>> 
>>>>>>>>>> MPI_T_ERR_CANTINIT --> MPI_T_ERR_CANNOTINIT (we do not use
>>>>>>>>>> abbrevations)
>>>>>>>>>> 
>>>>>>>>>> or
>>>>>>>>>> 
>>>>>>>>>> MPI_T_ERR_CANTINIT --> MPI_T_ERR_CANNOT_INIT (we do not
>> use
>>>>>>>>>> abbrevations)
>>>>>>>>>> 
>>>>>>>>>> In most cases, we do not combine words without underscore.
>>>>>>>>>> 
>>>>>>>>>> As far as I see, you can do the underscores.
>>>>>>>>>> Be sure to stay at maximum with 30 characters.
>>>>>>>>>> 
>>>>>>>>>> Current longest MPI_T constants:
>>>>>>>>>> 
>>>>>>>>>> 123456789012345678901234567890
>>>>>>>>>> MPI_T_PVAR_CLASS_HIGHWATERMARK
>>>>>>>>>> MPI_T_PVAR_CLASS_LOWWATERMARK
>>>>>>>>>> MPI_T_VERBOSITY_MPIDEV_DETAIL
>>>>>>>>>> 
>>>>>>>>>> Based on this, I would stay with you names based on your MPI_T
>>>>>>>>>> rule:
>>>>>>>>>> MPI_T_<structual-areas-with-underscores>_<final-name-without-
>> underscores>
>>>>>>>>>> 
>>>>>>>>>> i.e., I would only change
>>>>>>>>>> 
>>>>>>>>>> MPI_T_ERR_CANTINIT --> MPI_T_ERR_CANNOTINIT
>>>>>>>>>> 
>>>>>>>>>> Please tell me also your decision
>>>>>>>>>> that I can do the correct changes in chap-appLang
>>>>>>>>>> and you in chap-tools.
>>>>>>>>>> 
>>>>>>>>>> Best regards
>>>>>>>>>> Rolf
>>>>>>>>>> 
>>>>>>>>>> 
>>>>>>>>>> --
>>>>>>>>>> Dr. Rolf Rabenseifner . . . . . . . . . .. email
>>>>>>>>>> rabenseifner at hlrs.de
>>>>>>>>>> High Performance Computing Center (HLRS) . phone
>>>>>>>>>> ++49(0)711/685-65530
>>>>>>>>>> University of Stuttgart . . . . . . . . .. fax ++49(0)711 /
>>>>>>>>>> 685-65832
>>>>>>>>>> Head of Dpmt Parallel Computing . . .
>>>>>>>>>> www.hlrs.de/people/rabenseifner
>>>>>>>>>> Nobelstr. 19, D-70550 Stuttgart, Germany . (Office: Allmandring
>>>>>>>>>> 30)
>>>>>>>>>> 
>>>>>>>>> _______________________________________________
>>>>>>>>> mpi-forum mailing list
>>>>>>>>> mpi-forum at lists.mpi-forum.org
>>>>>>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>>>>>>>>> 
>>>>>>> 
>>>>>>> --
>>>>>>> Dr. Rolf Rabenseifner . . . . . . . . . .. email
>>>>>>> rabenseifner at hlrs.de
>>>>>>> High Performance Computing Center (HLRS) . phone
>>>>>>> ++49(0)711/685-65530
>>>>>>> University of Stuttgart . . . . . . . . .. fax ++49(0)711 /
>>>>>>> 685-65832
>>>>>>> Head of Dpmt Parallel Computing . . .
>>>>>>> www.hlrs.de/people/rabenseifner
>>>>>>> Nobelstr. 19, D-70550 Stuttgart, Germany . (Office: Allmandring 30)
>>>>>>> _______________________________________________
>>>>>>> mpi-forum mailing list
>>>>>>> mpi-forum at lists.mpi-forum.org
>>>>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>>>>>> 
>>>>>> 
>> ______________________________________________________________________
>> __
>>>>>> Martin Schulz, schulzm at llnl.gov, http://people.llnl.gov/schulzm
>>>>>> CASC @ Lawrence Livermore National Laboratory, Livermore, USA
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>> _______________________________________________
>>>>>> mpi-forum mailing list
>>>>>> mpi-forum at lists.mpi-forum.org
>>>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>>>>> 
>>>>> --
>>>>> Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner at hlrs.de
>>>>> High Performance Computing Center (HLRS) . phone ++49(0)711/685-
>> 65530
>>>>> University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832
>>>>> Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner
>>>>> Nobelstr. 19, D-70550 Stuttgart, Germany . (Office: Allmandring 30)
>>>> 
>>>> 
>> ______________________________________________________________________
>> __
>>>> Martin Schulz, schulzm at llnl.gov, http://people.llnl.gov/schulzm
>>>> CASC @ Lawrence Livermore National Laboratory, Livermore, USA
>>>> 
>>>> 
>>>> 
>>>> 
>>>> _______________________________________________
>>>> mpi-forum mailing list
>>>> mpi-forum at lists.mpi-forum.org
>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>>> 
>>> 
>>> _______________________________________________
>>> mpi-forum mailing list
>>> mpi-forum at lists.mpi-forum.org
>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>>> 
>> _______________________________________________
>> mpi-forum mailing list
>> mpi-forum at lists.mpi-forum.org
>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
> 
> _______________________________________________
> mpi-forum mailing list
> mpi-forum at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

________________________________________________________________________
Martin Schulz, schulzm at llnl.gov, http://people.llnl.gov/schulzm
CASC @ Lawrence Livermore National Laboratory, Livermore, USA

More information about the mpi-forum mailing list