Init/Finalize Clarification #37

jdinan · 2017-01-23T15:50:59Z

Clarify that shmem_init must be matched with a call to shmem_finalize.

Remove passage from API notes that contradicts the following semantic from the API description of shmem_init: "At the end of the OpenSHMEM program which it initialized, the call to shmem_init must be matched with a call to shmem_finalize." Signed-off-by: James Dinan <james.dinan@intel.com>

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan · 2017-02-22T16:17:05Z

Remove the following from finalize notes:

There is an implicit shmem_finalize at the end of main, so that having an
explicit call to shmem_finalize is optional. However, an explicit shmem_finalize may be required as an
entry point for wrappers used by profiling or other tools that need to perform their own finalization.

jdinan · 2017-02-22T16:17:19Z

Update mention of shmem_finalize in Annex A

jdinan · 2017-02-22T16:19:10Z

content/backmatter.tex

@@ -440,6 +440,10 @@ \section{Version 1.5}

 \begin{itemize}
 %
+\item Clarified that \FUNC{shmem\_init} must be matched with a call to
+    \FUNC{shmem\_finalize}.
+\\See Section \ref{subsec:shmem_init}.


Also ref shmem_finalize

Signed-off-by: James Dinan <james.dinan@intel.com>

naveen-rn · 2017-03-20T19:38:47Z

I'm bit confused with this change...Just to be sure, in this change we are not removing implicit finalize from the spec?

jdinan · 2017-03-20T19:48:44Z

This change removes implicit finalize in cases where the library was initialized with a call to shmem_init. We don't document this adequately (probably should be in deprecated interfaces), but implicit finalize can still be used with start_pes.

naveen-rn · 2017-03-20T20:03:51Z

At present, we have information about "implicit finalize" at two places - at shmem_init and shmem_finalize explanations. With this change, we seem to remove this information at both these places. At least, to me it looks like we have removed implicit finalize completely from the spec.
Can you please verify?

jdinan · 2017-03-20T20:51:11Z

Do you have an objection to the semantic that I mentioned in my previous comment, or are you requesting that we update the deprecated interfaces section to include implicit finalize as part of this proposal? Or something else?

I don't mind updating the deprecated interfaces documentation, including the start_pes section before we ratify this proposal.

naveen-rn · 2017-03-20T21:14:31Z

We are removing implicit finalize usage along with shmem_init in this PR. I don't want this semantics change without fixing the backward incompatibility issues.

This change removes implicit finalize in cases where the library was initialized with a call to shmem_init

At present, we require every shmem_init to be matched with shmem_finalize. But, we don't specifically say whether the finalize comes from explicit or implicit call. But this PR tends to make the explicit shmem_finalize usage a required model.

I like the intention behind this PR and this clarification is also important. So let us add new routines: shmem_finalized and shmem_initialized like MPI_initialized and MPI_finalized along with this change and completely remove implicit finalize from the specification.

naveen-rn · 2017-03-22T18:16:44Z

FWIU, as per the current specification the following is the desired and allowed usage model.

Use shmem_init along with explicit shmem_finalize and there is no implicit finalize at the end
Deprecated, but allow using start_pes() along with implicit finalize.

As you said, the current spec is inadequate with the following unanswered question:

Is using shmem_init() with implicit finalize a valid usecase? (May not be a desired model but is it valid?)
Is implicit finalize supposed to work only with start_pes() and a shmem_init() should always be used with shmem_finalize()?

This PR tries to clarify the above question, by removing details about implicit finalize from all places w.r.t to shmem_init. To enhance this clarification, as you suggested we should update the deprecated interfaces documentation, by including the start_pes section and also explicitly telling implicit finalize is valid only with start_pes.

naveen-rn · 2017-03-22T18:24:33Z

I take back my previous request, I don't have any use for implicit finalize now ;)
Assuming no one comes with the request for implicit finalize, we could proceed adding the reference-count semantics (either first-call or nested-call) to further increase the usability for long-term.
As we are dealing with almost similar changes and based on usage requirement priority,

We could either add this new semantics as part of this PR or
We could open two different tickets.

jdinan · 2017-03-23T13:53:49Z

Reference counting requires the library to support re-initialization, because of applications that may use more than one SHMEM library:

FOO_init(); /* e.g. I/O Library */
FOO_finalize();
BAR_init(); /* e.g. Compute Library */
BAR_finalize();

I suggest that we split these up -- use this ticket to clarify the current semantics as you described above and start a second ticket for reference counting. Would you be ok with that approach?

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan · 2017-04-10T17:58:01Z

@naveen-rn Could you please review the updated draft?
PDF version: main_spec.pdf

naveen-rn · 2017-04-11T19:40:02Z

Looks fine to me.

jdinan · 2017-04-28T18:50:37Z

content/execution_model.tex

-resources without terminating the program. Calling any \openshmem routine after
-\FUNC{shmem\_finalize} leads to undefined behavior.
+program concludes its use of the \openshmem library when all \acp{PE} call
+\FUNC{shmem\_finalize} or when any \ac{PE} calls \FUNC{shmem\_global\_exit}.


From @shamisp - Do we want to preserve the bit about completing all pending communication and releasing all resources in finalize?

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan · 2017-05-10T00:45:11Z

Special ballot changes: 53843d4

jdinan · 2017-05-25T16:28:33Z

content/execution_model.tex

-program finishes execution by returning from the main routine or when any PE
-calls \FUNC{shmem\_global\_exit}. When returning from main, \openshmem must
+program concludes its use of the \openshmem library when all \acp{PE} call
+\FUNC{shmem\_finalize} or when any \ac{PE} calls \FUNC{shmem\_global\_exit}.


Split into two sentences: "In addition, the program can be aborted when any PE calls shmem_global_exit."

BryantLam · 2017-05-25T16:30:49Z

content/execution_model.tex

-calls \FUNC{shmem\_global\_exit}. When returning from main, \openshmem must
+program concludes its use of the \openshmem library when all \acp{PE} call
+\FUNC{shmem\_finalize} or when any \ac{PE} calls \FUNC{shmem\_global\_exit}.
+During finalization, \openshmem must


Suggested

"finalization" -> "shmem_finalize"

"openshmem" -> "the openshmem library"

Fixed in c1dfa16, please take a look!

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan · 2017-05-31T18:44:30Z

For next time, special ballot readings: 53843d4..c1dfa16

Return values and error checking

James Dinan added 2 commits January 23, 2017 10:49

Add changelog entry for shmem_init clarification

09d3b9f

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan added the PendingReading label Feb 3, 2017

jdinan self-assigned this Feb 6, 2017

jdinan commented Feb 22, 2017

View reviewed changes

James Dinan added 3 commits March 8, 2017 10:48

Update changelog entry to ref shmem_finalize

1724797

Signed-off-by: James Dinan <james.dinan@intel.com>

Update Annex A reference to finalize

cbd1a26

Signed-off-by: James Dinan <james.dinan@intel.com>

Remove implicit finalize text from API notes

b4590fb

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan added HadReading PendingBallot SpecialBallot and removed PendingReading labels Mar 8, 2017

Merge branch 'osh_spec_next' into proposal/init-finalize

fb19471

James Dinan added 2 commits April 10, 2017 11:40

Merge branch 'osh_spec_next' into proposal/init-finalize

e2e172b

Add implicit finalize to start_pes and dep. annex

854efde

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan added PendingReading and removed HadReading PendingBallot labels Apr 12, 2017

jdinan removed the SpecialBallot label Apr 12, 2017

jdinan commented Apr 28, 2017

View reviewed changes

jdinan added HadReading and removed PendingReading labels Apr 28, 2017

Put back finalize description in exec model sec

53843d4

Signed-off-by: James Dinan <james.dinan@intel.com>

jdinan added PendingBallot SpecialBallot labels May 10, 2017

jdinan commented May 25, 2017

View reviewed changes

BryantLam reviewed May 25, 2017

View reviewed changes

Updates from 5/25/2017 reading

c1dfa16

Signed-off-by: James Dinan <james.dinan@intel.com>

manjugv added Ratified and removed HadReading PendingBallot SpecialBallot labels Jun 29, 2017

jdinan merged commit 5fa584e into openshmem-org:osh_spec_next Aug 1, 2017

BryantLam mentioned this pull request Aug 30, 2017

Deprecated "Implicit finalization" still supported? #124

Closed

jdinan deleted the proposal/init-finalize branch September 6, 2017 18:40

nspark pushed a commit to nspark/specification that referenced this pull request Sep 28, 2018

Merge pull request openshmem-org#37 from gmegan/issue_25+27

6f5982d

Return values and error checking

davidozog pushed a commit to davidozog/openshmem-specification that referenced this pull request Feb 20, 2019

Merge pull request openshmem-org#37 from gmegan/issue_25+27

cec3023

Return values and error checking

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Init/Finalize Clarification #37

Init/Finalize Clarification #37

jdinan commented Jan 23, 2017

jdinan commented Feb 22, 2017

jdinan commented Feb 22, 2017

jdinan Feb 22, 2017

naveen-rn commented Mar 20, 2017

jdinan commented Mar 20, 2017

naveen-rn commented Mar 20, 2017

jdinan commented Mar 20, 2017

naveen-rn commented Mar 20, 2017 •

edited

Loading

naveen-rn commented Mar 22, 2017

naveen-rn commented Mar 22, 2017 •

edited

Loading

jdinan commented Mar 23, 2017

jdinan commented Apr 10, 2017

naveen-rn commented Apr 11, 2017

jdinan Apr 28, 2017

jdinan May 10, 2017

jdinan commented May 10, 2017

jdinan May 25, 2017

BryantLam May 25, 2017 •

edited

Loading

jdinan May 25, 2017

jdinan commented May 31, 2017

Init/Finalize Clarification #37

Init/Finalize Clarification #37

Conversation

jdinan commented Jan 23, 2017

jdinan commented Feb 22, 2017

jdinan commented Feb 22, 2017

jdinan Feb 22, 2017

Choose a reason for hiding this comment

naveen-rn commented Mar 20, 2017

jdinan commented Mar 20, 2017

naveen-rn commented Mar 20, 2017

jdinan commented Mar 20, 2017

naveen-rn commented Mar 20, 2017 • edited Loading

naveen-rn commented Mar 22, 2017

naveen-rn commented Mar 22, 2017 • edited Loading

jdinan commented Mar 23, 2017

jdinan commented Apr 10, 2017

naveen-rn commented Apr 11, 2017

jdinan Apr 28, 2017

Choose a reason for hiding this comment

jdinan May 10, 2017

Choose a reason for hiding this comment

jdinan commented May 10, 2017

jdinan May 25, 2017

Choose a reason for hiding this comment

BryantLam May 25, 2017 • edited Loading

Choose a reason for hiding this comment

jdinan May 25, 2017

Choose a reason for hiding this comment

jdinan commented May 31, 2017

naveen-rn commented Mar 20, 2017 •

edited

Loading

naveen-rn commented Mar 22, 2017 •

edited

Loading

BryantLam May 25, 2017 •

edited

Loading