[Mpi-22] MPI_INIT assertions
Richard Treumann
treumann at [hidden]
Thu May 8 12:41:22 CDT 2008
Hi Alexander
First - changing the rules to let MPI_Info objects be created and
manipulated before MPI_INIT is much more intrusive on implementations. If
the MPI implementation uses some set of data structures to manage MPI
opaque objects, those data structures probably do not exist before
MPI_INIT. Passing a bit vector into MPI_INIT does not require significant
reorganization of the MPI implementation but allowing MPI_Info creation and
manipulation does. The concept of an "assertions" bit vector is not new for
MPI (see last paragreph) so using it does not require inventing two
mechanisms.
You have found a couple examples that flirt with violating the concept of a
"hint".
For soft spawn I would argue that any soft spawn call that would have
succeeded as a hard spawn must also succeed as a soft spawn no matter what
hint you provide. That is, if you try to spawn with maxprocs=16 and you
get 16 task your hint that you would take 8 has not resulted in semantic
thin ice. If your hint said you take 32 tasks, MPI can give you 16 and
again no semantic thin ice. Even if you ask for 16 on the call but agree
to 8 with a hint and a hard spawn for 8 tasks "would have worked", libmpi
is free to ignore the hint and fail the call because it cannot provide the
16.
On wdir (+ host, arch, path, file) it gets fuzzy. An MPI implementation is
free to ignore any of these and the user is free to omit them and hope the
MPI implementation has reasonable defaults. It could be argued that the MPI
implementation has the option of ignoring a hint when it cannot honor it
and falling back to the default. That would fit the "hint" concept but I
do not know if that was envisioned when these were put in MPI-2.
Actually we may want to look at doing a careful catagorization of the kinds
of inputs that an MPI_Info can convey. I may be making assumptions that are
not justified by a strict reading of the standard when I imply that
MPI_Info (key, value) pairs == hints. As part of MPI 3 we may want to
specify that the MPI_Info object can carry several kinds if "info" to an
MPI implementation after MPI_INIT.
hints - can never break semantic or trigger error messages
directives - do not break semantic but can make the difference between a
run succeeding and failing (e.g.the named wdir does not exist so libmpi
takes an error trying to use it)
assertions - can alter semantic. If we want a way to put assertions on
communicators for example I see no reason we could not use an Info object
to pass them in
more?
In MPI_FILE_OPEN we have MPI_MODE_UNIQUE_OPEN as a mode flag and say that
if the user says MPI_MODE_UNIQUE_OPEN and is wrong, his results cannot be
ensured. Silent corruption of the file is possible. We do not use the word
"assertion" though maybe we should have. In MPI 1sided synch operations we
have an integer IN argument we call an "assertion" and it can take flages
like MPI_MODE_NOCHECK, MPI_MODE_NOSTORE or MPI_MODE_NOPUT. We tell the
user that if he provides an assertion that is invalid he cannot assume the
program will work even though it would have without the assertion.
Dick
Dick Treumann - MPI Team/TCEM
IBM Systems & Technology Group
Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
Tele (845) 433-7846 Fax (845) 433-8363
"Supalov,
Alexander"
<alexander.supalo To
v_at_[hidden]> "MPI 2.2"
Sent by: <mpi-22_at_[hidden]>
mpi-22-bounces_at_li cc
sts.mpi-forum.org
Subject
Re: [Mpi-22] MPI_INIT assertions
05/08/2008 10:52
AM
Please respond to
"MPI 2.2"
<mpi-22_at_lists.mpi
-forum.org>
Dear Dick,
Thank you, I think I start feeling the difference between the assertions
and the hints. To be completely sure, where would you put the soft spawning
info key: is this an assertion or a hint? What about other standardized
info elements that materially influence the program execution or may even
predetermine its success or failure (like wdir)? More generally, will the
difference between assertions and hints justify the use of two different
mechanisms to pass the data across the MPI boundary?
Here's I think what we can save by dropping certain MPI features:
Dynamic process support: less overhead in the progress engine, easier
global rank handling.
File I/O: smaller requests, easier wait/test functions.
One-sided ops: no passive target w/o MPI calls - no extra progress
thread.
Communicator & group management: better memory footprint.
Non-blocking communication: easier ordering, simplified request
handling.
Heterogeneity: better memory footprint, easier data handling.
Derived datatypes (especially those with holes): better memory
footprint.
Message tagging: better support for stable dataflow exchanges,
smaller packets.
Best regards.
Alexander
From: mpi-22-bounces_at_[hidden]
[mailto:mpi-22-bounces_at_[hidden]] On Behalf Of Richard Treumann
Sent: Tuesday, May 06, 2008 5:21 PM
To: MPI 2.2
Subject: Re: [Mpi-22] MPI_INIT assertions
In my view, each assertion we define in the standard should be justified by
a good explanation of how it allows an MPI implementation to be more
efficient and what cost/risk it adds In effect, a cost/benefit analysis. No
plausible rationale - no assertion.
An assertion is a statement about the application. It is not a request for
libmpi to "provide something". In my proposal, an application that
correctly uses some assertion must still be written to work on an MPI
implementation that ignores that assertion. An application that incorrectly
uses some assertion can fail or give wrong answers. You can picture an
assertion as the application author granting libmpi permission to "NOT
provide something" at the MPI implementors discretion.
In MPI 2 we maintained a careful distinction between hints and assertions.
The MPI implementation is not allowed to give wrong results or error
messages when an application provides an incorrect hint. If an application
could provide a hint that it would not use MPI_CANCEL on a particular
communicator and then called MPI_CANCEL, it would be OK for libmpi to use a
really slow CANCEL protocol but not OK for it to issue a fatal error or
ignore the CANCEL. That means that the freedom to "exploit" a hint in
libmpi is very limited and less care is required of the user in providing
hints. Much of what is being proposed should be treated as hints and could
use a different and extensible mechanism. We can consider extending the
MPI_Info usage in MPI-3 to address information passing that is hint-like.
The MPI standard intends applications to be platform agnostic and source
level portable and we should not break that. An assertions interface for
making statements about the application will not limit this kind of
portability. If there are no MPI_CANCEL calls today, there are no
MPI_CANCEL calls after recompilation for a new platform. If we add features
for platform tuning (like how many tasks per node or eager threshhold) we
could damage portability.
In my first outline, I threw in MPI-IO and MPI-1SC without any rationale
for how they are useful. You added: "No communicator & group management",
"No non-blocking communication", "No message tagging" and "No derived
datatypes". It is not clear to me they are all useful so a decent rationale
would be needed for each. That makes 6 bits that may be free.
The thread support level parameter in MPI_INIT_THREAD is a request to the
MPI library to provide some level of thread support and I question how
cleanly that maps into the "assertions" model. The assertions model may
only require one bit for "MPI_NO_THREAD_CONTENTION". That frees 3 more of
the bits you suggest are already taken. If we conclude that all 4 thread
support options are really useful it may be best to add the assertions flag
to MPI_INIT_THREAD and keep the thread support controls as they are. Even
the MPI_NO_THREAD_CONTENTION bit frees up then. My original thought was
that we would add a new function that adds a parameter to MPI_INIT_THREAD.
I am convinced that the number of assertions defined by the MPI standard
should be very limited because each brings some risk and complicates life
for library writers. If we end up with as many as 32 assertions I will
apologize for having offered the idea in the first place and probably
advocate a no vote. If MPI 2.2 provided only MPI_NO_EAGER_THROTTLE,
MPI_NO_SEND_CANCEL, MPI_NO_REQUEST_MIX, MPI_NO_ANY_SOURCE and
MPI_NO_REDUCTION_ORDER I would be content and debate about adding more
could wait for MPI 3.0. The assertion that the application does not try to
match persistent send/recv with nonpersistent recv/send would be a nice 2.2
bonus. That accounts for 6 bits.
The query function could be something as simple as:
MPI_Query_assertions(int *stated, int *exploited)
The author of libfred could code in his Init_fred routine:
int stated_by_app, exploited_by_lib;
MPI_Query_assertions( &stated_by_app, &exploited_by_lib)
if (exploited_by_lib & MPI_NO_SEND_CANCEL)
fatal_error("Assertion MPI_NO_SEND_CANCEL is not compatable with Fred
Lib")
else if (stated_by_app & MPI_NO_SEND_CANCEL)
warn("Assertion MPI_NO_SEND_CANCEL is not compatable with Fred Lib.
It is not exploited by your MPI but your application will not port to an
MPI that does exploit MPI_NO_SEND_CANCEL");
If Fred is really cautious he can code
If (stated_by_app != 0) fatal_error("Fred does not like assertions")
and see if anybody asks for some assertion to be tolerated in the next
release of FredLib.
You mention "No eager buffering" which is different than what I proposed
and is not logically an "assertion". MPI applications should not depend
semantically on eager buffering today and the standard is clear that a MPI
implementation which has no eager buffering can still be 100% compliant.
The standard requires that IF an MPI implementation provides an eager
protocol as a performance enhancement, it must also provide a safety
feature which can be expensive. Many (perhaps most) applications can
benefit from an eager protocol but do not need the safety feature because
the application is self regulating. These applications could assert: "I do
not require the safety feature." which would allow them the benefit of
eager protocol without the costs of the safety feature.
I can see a rationale for some kind of hints related to eager protocol
buffer space and threshhold. There are many other kinds of granular "hint"
support that may be useful but all that should be kept distinct from this
proposal (In my opinion).
Regards - Dick
Dick Treumann - MPI Team/TCEM
IBM Systems & Technology Group
Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
Tele (845) 433-7846 Fax (845) 433-8363
mpi-22-bounces_at_[hidden] wrote on 05/06/2008 07:26:03 AM:
> Dear Dick,
>
> Thanks. What about passing more verbose information to and from an
> MPI implementation? Say, eager threshold value, if it's meaningful?
> There are many less controversial examples, for example, to let the
> caller know how many processes are placed onto the node, etc. Simple
> assertions won't work in this case.
>
> Back to assertions. I thought we might consider
> MPI_Initialized_subset(*flag,required,*provided) or something.
> Setting "required" to 0 would return the set of provided assertions
> in "provided". Setting "required" to something else will also set
> the "flag" to true or false, depending on whether there's a match.
>
> The special MPI_Init_subset(*argc,***argv,required,*provided) would
> complement this.
>
> Regarding 32 bits, let's count (I'm taking the joint list from the
> subset proposal that includes your earlier points):
>
> 0,1,2,3 - MPI_THREAD_SINGLE, FUNNELED, SERIALIZED, MULTIPLE
> 0x00000004 - No dynamic process support
> 0x00000008 - No file I/O
> 0x00000010 - No one-sided ops
> 0x00000020 - No communicator & group management
> 0x00000040 - No non-blocking communication
> 0x00000080 - No message cancellation
> 0x00000100 - Persistent ops on both sides
> 0x00000200 - No heterogeneity
> 0x00000400 - No derived datatypes (especially those with holes)
> 0x00000800 - No MPI_ANY_SOURCE
> 0x00001000 - No message tagging
> 0x00002000 - No reduction order
> 0x00004000 - No eager buffering
> 0x00008000 - No mixed request types in wait/test
>
> We have 16 bits left. Note that there may be some predefined masks
> for well defined combinations of the above (e.g., MPI-1 = no dynamic
> processes, no file I/O, no threads, etc.).
>
> Best regards.
>
> Alexander
>
> From: mpi-22-bounces_at_[hidden] [mailto:mpi-22-
> bounces_at_[hidden]] On Behalf Of Richard Treumann
> Sent: Monday, May 05, 2008 11:53 PM
> To: MPI 2.2
> Subject: [Mpi-22] MPI_INIT assertions
> I changed the subject to be meaningful - it had been "Re: [Mpi-22]
> mpi-22 Digest, Vol 2, Issue 7"
>
> I do think this proposal is within the scope of the MPI 2.2 rules so
> I am pleased to hear this is being considered.
>
> I am 100% convinced that a query function should be part of this
> proposal, probably with 2 query flavors. One flavor of the query
> would respond with the set of assertions the application
> MPI_INIT_xxx call had provided and another would respond with the
> set of assertions the MPI implementation is actually exploiting.
>
> One library author may decide he will use if/else logic based on how
> the MPI implementation will behave. The application may assert
> MPI_NO_yyy but if the MPI implementation does not exploit MPI_NO_yyy
> then the library can still use the code that depends on yyy. Another
> library author may depend unconditionally on yyy. On an MPI
> implementation that does not exploit MPI_NO_yyy he may want to issue
> a warning that assertion MPI_NO_yyy is non-portable but let the job
> run. On one that does exploit MPI_NO_yyy he would abort the job. Or,
> he may decide to simply abort any job that asserts MPI_NO_yyy to
> avoid having the library suddenly quit working when the customer
> upgrades to an MPI implementation that does exploit MPI_NO_yyy.
>
> I think 32 assertions is probably more than enough
>
> Even a handful of defined assertions can raise testing costs. For
> simple cases like MPI_NO_ANY_SOURCE it is easy for a user to judge
> whether a piece of code is OK but if we begin to add subtle or
> narrow semantic assertions it gets harder. Some library providers
> will be tempted to say "I cannot proof read and test my code to a
> degree that will allow me to accept 28 specific assertions and
> forbid 4. I will simply forbid every subtle assertion I cannot
> afford to test."
>
> I predict that for large applications developed by teams and for
> community open source efforts the design leads will consider
> requiring that all parts be written to live within a few carefully
> chosen assertions. The design leads will not want to provide a list
> of 26 subtle or narrow assertions and require that everyone respect
> all 26 in the code they contribute.
>
> Many distinct assertions also could become a big test cost for MPI
> implementors and customers who must qualify an MPI implementation
> before trusting their business to it. If there were 64 or more
> assertions, how would a tester decide what combinations of
> assertions must be tested and then create suitable test cases?
>
> Dick
>
> Dick Treumann - MPI Team/TCEM
> IBM Systems & Technology Group
> Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
> Tele (845) 433-7846 Fax (845) 433-8363
>
>
> mpi-22-bounces_at_[hidden] wrote on 05/05/2008 01:28:06 PM:
>
> > Dear Dick,
> >
> > Thank you. We can actually introduce what you propose, possibly with
> > a query function to make it still easier to live with, as early as
> > in MPI 2.2, as a subset precursor. Judging by the discussion in
> > Chicago, subsets may not need much more than that in the end,
> > possibly with a little more flags and semantics added in MPI-3.
> >
> > The reservation against 32- (or for that matter, 64-) bit limitation
> > is the only one I have at the moment. Not being able to attach
> > assertions to communicators, etc. may be missed by some advanced
> > programmers, but here we need to be pragmatic: who will ever want to
> > go that deep?
> >
> > Best regards.
> >
> > Alexander
> >
> > From: mpi-22-bounces_at_[hidden] [mailto:mpi-22-
> > bounces_at_[hidden]] On Behalf Of Richard Treumann
> > Sent: Monday, May 05, 2008 7:18 PM
> > To: MPI 2.2
> > Subject: Re: [Mpi-22] mpi-22 Digest, Vol 2, Issue 7
>
> > Hi Alexander
> >
> > I have no objection to citing my "assertions" proposal in the
> > subsetting discussions. I do want to keep it clear that this
> > proposal is intended to be as simple as practical to implement, exploit
and
> > live with.
> >
> > "Live with" applies to 3rd party library authors or anyone else who
> > must write MPI code but does not know and control the structure of
> > the entire application. That guy must "live with" the decisions made
> > by whoever coded the MPI_INIT piece. "Live with" also applies to
> > whoever must test or certify a specific MPI implementation.
> >
> > Thanks - Dick
> >
> > Dick Treumann - MPI Team/TCEM
> > IBM Systems & Technology Group
> > Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
> > Tele (845) 433-7846 Fax (845) 433-8363
> >
> > [image removed] "Supalov, Alexander" <alexander.supalov_at_[hidden]>
> >
>
> >
> > "Supalov, Alexander" <alexander.supalov_at_[hidden]>
> > Sent by: mpi-22-bounces_at_[hidden]
> > 04/26/2008 04:03 AM
> >
> > Please respond to
> > "MPI 2.2" <mpi-22_at_[hidden]>
> >
> > [image removed]
> > To
> >
> > [image removed]
> > "MPI 2.2" <mpi-22_at_[hidden]>
> >
> > [image removed]
> > cc
> >
> > [image removed]
> >
> > [image removed]
> > Subject
> >
> > [image removed]
> > Re: [Mpi-22] mpi-22 Digest, Vol 2, Issue 7
> >
> > [image removed]
> >
> > [image removed]
> >
> >
> > Dear Dick,
> >
> > Thank you. Would you mind if I cite your proposal in the subsets
> > discussion? Yours looks like a good alternative to the thinking of
> > some of us that subsets might be very rich and mutable, and to
> > Jeff's proposal on hints I've already cited there with his permission.
> >
> > Best regards.
> >
> > Alexander
> >
> > From: mpi-22-bounces_at_[hidden] [mailto:mpi-22-
> > bounces_at_[hidden]] On Behalf Of Richard Treumann
> > Sent: Thursday, April 24, 2008 6:16 PM
> > To: MPI 2.2
> > Subject: Re: [Mpi-22] mpi-22 Digest, Vol 2, Issue 7
> > Dick Treumann - MPI Team/TCEM
> > IBM Systems & Technology Group
> > Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
> > Tele (845) 433-7846 Fax (845) 433-8363
> >
> >
> > mpi-22-bounces_at_[hidden] wrote on 04/24/2008 11:33:42 AM:
> >
> > > Hi,
> > >
> > > Note that this is an argument for making the assertions optional:
those
> > > who don't care don't have to use them. Those who care should use them
> > > correctly or else. As usual.
> > >
> > > Best regards.
> > >
> > > Alexander
> > >
> >
> > Hi Alexander
> >
> > The assertions are optional in this proposal. If this is added to
> > the MPI standard the minimal impacts (day one impacts) are:
> >
> > ==
> > To application writers (none) - MPI_INIT and MPI_INIT_THREAD still
> > work. MPI_INIT_THREAD_xxx can be
> > passed 0 (zero) as the assertions bit vector.
> >
> > To MPI Implementors (small) - subroutine MPI_INIT_THREAD_xxx can be
> > a clone of MPI_INIT_THREAD under the covers. If the Forum decides
> > the query function is for asking what assertions are being honored,
> > the implementation can just return "none" to every query. If there
> > is also a query for what assertions have been made then there are a
> > few more lines of code the implementor must write to preserve the
> > value so it can be returned(maybe 10 lines)
> >
> > Writers of opaque libraries (small) - call the query function at
> > library init time and if any assertions are found, issue an error
> > message and kill the job. This is awkward for a library that wants
> > to support every MPI whether it has implemented the new query
> function or not.
> > ==
> >
> > As MPI implementations begin to take advantage of assertions there
> > is more work for the MPI implementor and the library author must
> > begin to think about whether his customer will be upset if the
> > library simply outlaws all assertions.
> >
> > The library author will never be wrong if he simply forbids
> > assertions forever. If they become valuable he will feel the
> > pressure to work it out.
> >
> > The MPI implementor will never be wrong if he adds the API but
> > simply ignores assertions forever. If they become valuable he will
> > feel the pressure to honor some at least.
> > ---------------------------------------------------------------------
> > Intel GmbH
> > Dornacher Strasse 1
> > 85622 Feldkirchen/Muenchen Germany
> > Sitz der Gesellschaft: Feldkirchen bei Muenchen
> > Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer
> > Registergericht: Muenchen HRB 47456 Ust.-IdNr.
> > VAT Registration No.: DE129385895
> > Citibank Frankfurt (BLZ 502 109 00) 600119052
> >
> > This e-mail and any attachments may contain confidential material for
> > the sole use of the intended recipient(s). Any review or distribution
> > by others is strictly prohibited. If you are not the intended
> > recipient, please contact the sender and delete all copies.
> > _______________________________________________
> > mpi-22 mailing list
> > mpi-22_at_[hidden]
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-22
> > ---------------------------------------------------------------------
> > Intel GmbH
> > Dornacher Strasse 1
> > 85622 Feldkirchen/Muenchen Germany
> > Sitz der Gesellschaft: Feldkirchen bei Muenchen
> > Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer
> > Registergericht: Muenchen HRB 47456 Ust.-IdNr.
> > VAT Registration No.: DE129385895
> > Citibank Frankfurt (BLZ 502 109 00) 600119052
> >
> > This e-mail and any attachments may contain confidential material for
> > the sole use of the intended recipient(s). Any review or distribution
> > by others is strictly prohibited. If you are not the intended
> > recipient, please contact the sender and delete all copies.
> > _______________________________________________
> > mpi-22 mailing list
> > mpi-22_at_[hidden]
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-22
> ---------------------------------------------------------------------
> Intel GmbH
> Dornacher Strasse 1
> 85622 Feldkirchen/Muenchen Germany
> Sitz der Gesellschaft: Feldkirchen bei Muenchen
> Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer
> Registergericht: Muenchen HRB 47456 Ust.-IdNr.
> VAT Registration No.: DE129385895
> Citibank Frankfurt (BLZ 502 109 00) 600119052
>
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.
> _______________________________________________
> mpi-22 mailing list
> mpi-22_at_[hidden]
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-22
---------------------------------------------------------------------
Intel GmbH
Dornacher Strasse 1
85622 Feldkirchen/Muenchen Germany
Sitz der Gesellschaft: Feldkirchen bei Muenchen
Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer
Registergericht: Muenchen HRB 47456 Ust.-IdNr.
VAT Registration No.: DE129385895
Citibank Frankfurt (BLZ 502 109 00) 600119052
This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
_______________________________________________
mpi-22 mailing list
mpi-22_at_[hidden]
http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-22
*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpi-22/attachments/20080508/12be2f50/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: graycol.gif
Type: application/octet-stream
Size: 156 bytes
Desc: not available
URL: <http://lists.mpi-forum.org/pipermail/mpi-22/attachments/20080508/12be2f50/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: pic25566.gif
Type: application/octet-stream
Size: 1255 bytes
Desc: not available
URL: <http://lists.mpi-forum.org/pipermail/mpi-22/attachments/20080508/12be2f50/attachment-0001.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ecblank.gif
Type: application/octet-stream
Size: 45 bytes
Desc: not available
URL: <http://lists.mpi-forum.org/pipermail/mpi-22/attachments/20080508/12be2f50/attachment-0002.obj>
More information about the Mpi-22
mailing list