[Openshmem-list] Effects of removing implicit finalize from the specification
Bob Cernohous
bcernohous at cray.com
Tue Mar 21 21:41:51 UTC 2017
In general, I think libraries with side effects like initializing/finalizing SHMEM are problematic. [It’s especially bad for something like dmapp_init which takes config parameter so neither FOO nor BAR can count on controlling the initialized configuration. Same thing, I guess, if FOO or BAR were trying to set env’s internally before init.]
But everything you say is true and if those are all the goals then reference counting is fine.
But there was, I believe, a customer request to support exit w/finalize. Ie. if a process took an early exit, not a shmem_global_exit, wait for other PEs to finalize/exit normally.
Or, imagine FOO and BAR didn’t explicitly have finalize functions but relied on atexit to finalize.
If you want atexit(shmem_finalize) (or the current implicit finalize) to wait and truly finalize, then it can’t just be decrementing a reference count and executing no-op.
From: Nicholas Park [mailto:nspark at computer.org]
Sent: Tuesday, March 21, 2017 4:20 PM
To: Naveen Ravichandrasekaran <nravi at cray.com>
Cc: Bob Cernohous <bcernohous at cray.com>; Kuehn, Jeff <jakuehn at lanl.gov>; Jeff Hammond <jeff.science at gmail.com>; openshmem-list at openshmem.org
Subject: Re: [Openshmem-list] Effects of removing implicit finalize from the specification
I'm very concerned about first-call semantics on shmem_finalize(). I think this effectively breaks general-purpose library support for OpenSHMEM.
Consider two libraries FOO and BAR that both build on OpenSHMEM.
Claim: Library interoperability has two goals:
1) Libraries FOO and BAR can both initialize and finalize the OpenSHMEM library
2) Users of FOO need not make OpenSHMEM calls can can interact solely with FOO
Under the proposed first-call semantic, Goal #1 is violated. A call sequence of:
FOO_init() --> includes call to shmem_init() --> actually initializes
BAR_init() --> includes call to shmem_init() --> no-op
BAR_finalize() --> includes call to shmem_finalize() --> actually finalizes
FOO_finalize() --> includes call to shmem_finalize() --> no-op
means that FOO_finalize() will fail if it calls any OpenSHMEM function other than shmem_finalize(). It could non-trivially require calls to shmem_free(), shmem_my_pe(), shmem_barrier_all().
In contrast, the reference-counted semantic leads to valid behavior:
FOO_init() --> includes call to shmem_init() --> actually initializes
BAR_init() --> includes call to shmem_init() --> no-op
BAR_finalize() --> includes call to shmem_finalize() --> no-op
FOO_finalize() --> includes call to shmem_finalize() --> actually finalizes
Currently, Goal #2 is violated by the fact that calling shmem_init() more than once is undefined. Each library must make the assumption that OpenSHMEM is either initialized/finalized solely by them or external to them. Only the latter is truly sane, but then means that a user interacting solely with FOO (without loss of generality) has to make OpenSHMEM calls (and may not even be aware that FOO relies on OpenSHMEM). I recognize we're trying to solve that in this discussion, but I wanted to make this goal explicit for library support.
Nick
On Tue, Mar 21, 2017 at 12:54 PM, Naveen Ravichandrasekaran <nravi at cray.com<mailto:nravi at cray.com>> wrote:
+1 for the first-call semantics – init/finalize on the first-call and then no-op after that.
-Naveen N Ravi.
From: Bob Cernohous
Sent: Tuesday, March 21, 2017 11:09 AM
To: Kuehn, Jeff <jakuehn at lanl.gov<mailto:jakuehn at lanl.gov>>; Naveen Ravichandrasekaran <nravi at cray.com<mailto:nravi at cray.com>>; Jeff Hammond <jeff.science at gmail.com<mailto:jeff.science at gmail.com>>
Cc: openshmem-list at openshmem.org<mailto:openshmem-list at openshmem.org>
Subject: RE: [Openshmem-list] Effects of removing implicit finalize from the specification
I would vote for first-call semantics. The email example wasn’t particularly interesting for atexit. The only reason to want atexit(shmem_finalize) is that there might be a code path that doesn’t explicitly shmem_finalize. I think you want atexit to finalize, not decrement a reference count. So init/finalize on first-call and no-op after that?
From: Openshmem-list [mailto:openshmem-list-bounces at openshmem.org] On Behalf Of Kuehn, Jeff
Sent: Tuesday, March 21, 2017 10:48 AM
To: Naveen Ravichandrasekaran <nravi at cray.com<mailto:nravi at cray.com>>; Jeff Hammond <jeff.science at gmail.com<mailto:jeff.science at gmail.com>>
Cc: openshmem-list at openshmem.org<mailto:openshmem-list at openshmem.org>
Subject: Re: [Openshmem-list] Effects of removing implicit finalize from the specification
I have a very small concern with nested semantics vs first call semantics in the case of init and finalize. If one considers the cases of an extra init or finalize and a missing init or finalize, it seems that the observable side effects of those bugs, are more local to the problem code, and thus easier to debug in the case of first-call semantics, and more non-local in the case of nested semantics. It’s a small concern, but what do the rest of you think?
Regards,
Jeff
On 3/20/17, 19:51, "Openshmem-list on behalf of Naveen Ravichandrasekaran" <openshmem-list-bounces at openshmem.org<mailto:openshmem-list-bounces at openshmem.org> on behalf of nravi at cray.com<mailto:nravi at cray.com>> wrote:
Hi Jeff,
> We either need to allow multiple calls to initialize/finalize or we
> need to have is_initialized/is_finalized queries. MPI does the latter
> but since it is utterly trivial to ref count in this case, we should just
> do better than MPI and do the former.
My requirement is not to support multiple shmem_finalize calls. I was just
merely stating the fact that the side-effect of implicit finalize in the current
specification allows the usage of shmem_finalize multiple times.
What we really need is a semantics which successfully performs finalize or
do no-op when the user uses atexit(shmem_finalize) in their application.
shmem_init(); /* ref_count++ -> actual initialization */
shmem_init(); /* ref_count++ -> no-op */
shmem_finalize(); /* ref_count-- -> no-op */
shmem_finalize(); /* ref_count-- -> actual finalization */
FWIU, the reference counter method would be useful for handling
multiple shmem_init and shmem_finalize calls, but it won’t probably
work for handling atexit(shmem_finalize).
In this example, it is unclear how the reference counter works.
int main(void) {
shmem_init(); /* ref_count++ -> actual initialization */
atexit(shmem_finalize);
shmem_finalize(); /* ref_count-- -> actual finalization */
return 0; /* unclear what should happen here */
}
Irrespective of the number of times shmem_init is used, we need to
successfully handle atexit(shmem_finalize). So, I would prefer adding
new APIs in SHMEM to query is_initialized/is_finalized and let the
users handle atexit(shmem_finalize) scenario correctly.
-Naveen N Ravi.
From: Jeff Hammond [mailto:jeff.science at gmail.com]
Sent: Monday, March 20, 2017 4:01 PM
To: Naveen Ravichandrasekaran <nravi at cray.com<mailto:nravi at cray.com>>
Cc: openshmem-list at openshmem.org<mailto:openshmem-list at openshmem.org>
Subject: Re: [Openshmem-list] Effects of removing implicit finalize from the specification
If we permit multiple calls to finalize then we should also permit multiple calls to initialize. This makes it easy on applications that use libraries and the libraries do not know whether they should call initialize or not.
We either need to allow multiple calls to initialize/finalize or we need to have is_initialized/is_finalized queries. MPI does the latter but since it is utterly trivial to ref count in this case, we should just do better than MPI and do the former.
We could add return codes on initialize/finalize to allow users to know whether the calls were no-ops or not. This does not break any existing code because users do not have to assign return codes to variables.
Jeff
On Mon, Mar 20, 2017 at 1:37 PM, Naveen Ravichandrasekaran <nravi at cray.com<mailto:nravi at cray.com>> wrote:
At present, the specification allows the multiple usage of shmem_finalize.
This seems to be a defined behavior and multiple subsequent calls to finalize
are no-ops.
The following example should pass as per the current specification;
Example:1
int main(void) {
shmem_init();
shmem_finalize();
shmem_finalize();
return 0;
}
In Example:1, we have three finalize (2 explicit + 1 implicit) operations.
Even any common SHMEM usage, will have two finalize operations
(1 explicit + 1 implicit). Even though the specification is not clear about the
above usage, this is the side-effect of having implicit finalize at exit() or at
return from main.
Example:2
int main(void) {
shmem_init();
atexit(shmem_finalize);
return 0;
}
During F2F meeting, we decided to drop implicit finalize in the specification.
The argument was that the users can explicitly call "atexit" in the application
to achieve the above behavior as shown in Example:2.
Example:3
int main(void) {
shmem_init();
atexit(shmem_finalize);
shmem_finalize();
return 0;
}
FWIU, if we remove implicit finalize from the specification - Example:2 is
guaranteed to work but the behavior on Example:3 is undefined.
To allow the explicit atexit usage, we either
1. Need to specify that multiple calls to finalize are no-ops or
2. We need to have a new API - "shmem_finalized()" so that the users can
have a wrapper from atexit which checks shmem_finalized() before calling
shmem_finalize().
Please let me know, if this analysis looks valid.
-Naveen N Ravi
Cray Inc.
_______________________________________________
Openshmem-list mailing list
Openshmem-list at openshmem.org<mailto:Openshmem-list at openshmem.org>
http://www.openshmem.org/mailman/listinfo/openshmem-list
--
Jeff Hammond
jeff.science at gmail.com<mailto:jeff.science at gmail.com>
http://jeffhammond.github.io/
_______________________________________________
Openshmem-list mailing list
Openshmem-list at openshmem.org<mailto:Openshmem-list at openshmem.org>
http://www.openshmem.org/mailman/listinfo/openshmem-list
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.openshmem.org/pipermail/openshmem-list/attachments/20170321/adc8b1d4/attachment-0001.html>
More information about the Openshmem-list
mailing list