[Mpi3-ft] Ticket 323 - status?
Richard Graham
richardg at mellanox.com
Thu May 31 12:55:22 CDT 2012
So, there was a point in time that we did invite people from the C/R community to the meetings, and did discuss these sort of things - mostly on the phone, as this was a direction that was also explored early on. The conclusion was that there is so little commonality between checkpointering mechanisms, that there was nothing to standardized on - the C/R community would first need to agree on some common approach, and Paul Hargrove who had suggested gathering those folks never got around to arranging for such discussions.
As I say, there is a long history here, and nothing was considered as taboo. The direction taken is one that works well when for a class of apps, and has been used by those apps that have used the FT-MPI approach, as well as the half dozen or so apps that have worked with the version of the proposal as it was in 2011, before collectives were dropped.
This is a complementary approach to C/R, one that gives apps a different tool in their tool kit, as Richard Barrett from Sandia said in the early phase of the proposal. This is also not orthogonal to C/R, I am sure there are cases where data will have to be read from some saved location , but this lets one start with a different configuration, and NOT loose data that is still good, allowing for local forms of recovery, which is NOT possible to do with the current form of C/R.
As I said, the group went through quite an extensive set of discussions over the past 4 years, it is just that those new to the process have not gone back to look through all the information on the wiki to learn this history, and as a w/g this is part of our collective knowledge, and we have not done well in reminding folks of this.
Also, let me remind folks that there are two more proposals coming up in this area - Tony's aspect based proposal, and LLNL's (Greg B. is the originator) for adding piggy back support. Both add yet more tools to deal with failure. I guess that my main objection to the points that have been raised is to the notion that "one size fits all" - it does not for programming approaches today, why would we think that it does for recovery under failure.
Rich
-----Original Message-----
From: mpi3-ft-bounces at lists.mpi-forum.org [mailto:mpi3-ft-bounces at lists.mpi-forum.org] On Behalf Of Sur, Sayantan
Sent: Friday, June 01, 2012 2:30 AM
To: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group
Subject: Re: [Mpi3-ft] Ticket 323 - status?
Ralph has provided some great insights about the unease of the rest of the forum to vote "yes" on this proposal.
It appears to me that there is still a research question out there that needs to be answered before this vote becomes uncontroversial. Given that checkpoint/restart can be made significantly faster with the advent of new memory technologies ... it is not a foregone conclusion that detecting faults on the fly, recovering and the pain involved in the process (both to MPI devs and app/lib devs) is worth it. Research needs to prove it in the modern context - and not by citing older results where checkpoints were taken to cluster FS based on HDD!
http://www.hpl.hp.com/people/naveen_muralimanohar/taco11.pdf
Looking back - part of me wishes that the WG had moved to improve C/R by providing apps a mechanism to indicate when is a good time to checkpoint. Then, there are libs like SCR that have been around and proven useful. Why can't interfaces like those be incorporated in MPI? There are other issues like tackling recovery for apps that do I/O (persistence makes it hard to recover).
Thanks.
===
Sayantan Sur, Ph.D.
Intel Corp.
> -----Original Message-----
> From: mpi3-ft-bounces at lists.mpi-forum.org [mailto:mpi3-ft-
> bounces at lists.mpi-forum.org] On Behalf Of Aurélien Bouteiller
> Sent: Thursday, May 31, 2012 8:33 AM
> To: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group
> Subject: Re: [Mpi3-ft] Ticket 323 - status?
>
> The proposal does not add significant complexity to the code base,
> especially when no fault tolerance is supported. I'm not waiving hands
> here, we have an implementation of this proposal that does not support
> fault tolerance, the diff is less than 1k lines total, mostly "stub"
> code for the interfaces that remaps to existing MPI constructs. On the
> matter of most interest to you, no changes are required in the runtime
> in this case (not a single line diff). Please have a look at the implementation yourself, it is available.
>
> Beside this proposal is not defining a recovery model, it is merely
> defining the error raised when failure happen and the following state
> of MPI, nothing more. Defining proper recovery models is left to
> user-level libraries, which benefit from this proposal by being
> portable across different MPI versions (even those that do not support
> effective fault tolerance because the target machine is stable, or the implementors were lazy).
>
> The large number of abstain votes proves that we didn't had time to
> convince people of the soundness of the approach, there was still
> significant uncertainty at the time of the vote in the mind of many
> participants. This is an important information, as it seems we often
> hear comments that are not founded on hard facts, but rather on
> feelings or fears (including that the text had been changed very close
> to the deadline, a very valid concern). We might want to make sure to
> publicize more the results we already have, and arguably, the tight
> schedule for 3.0 inclusion didn't left that much opportunity for this and has been detrimental.
>
> Aurelien
>
>
> Le 31 mai 2012 à 09:07, Ralph Castain a écrit :
>
> > Guess we can agree to disagree on your conclusions. It is true that
> > there
> are non-HPC users that want to run past errors (my own org included),
> but that problem is relatively easier to solve than the overall MPI
> issue as it doesn't involve things like collectives. As for HPC, I'll
> note that the OMPI mailing list has nearly zero questions for even the
> current level of FT, and those that do come in are from students doing
> research (not production users). And yes, I know bigger clusters are
> in the works, but as I noted, there are other solutions being worked for them too.
> >
> > My point was solely that there are multiple ways of addressing the
> problem, and that making major alterations to MPI is only one of them
> - and possibly a less attractive approach to some of the alternatives.
> So it may well be that people would prefer to leave MPI alone for now,
> pending more research that reduces the risk and performance penalties.
> >
> > In that vein, I would suggest you consider releasing an "FT-MPI"
> > again, or
> perhaps defining a quasi-standard set of MPI extensions that an
> implementation that wants to provide an FT version could support. The
> latter would still meet your desire to provide FT to a broader
> audience, but would avoid forcing ALL implementations to support
> something that their customers may not want.
> >
> > Let me re-emphasize: I'm not saying this was all for naught. I'm
> > trying to
> suggest reasons why the body might resist adopting this proposal, and
> possible paths forward that would still support the work. Forcing a
> community to adopt FT as "standard" may not be the best way in the
> near term.
> >
> > Ralph
> >
> >
> > On Wed, May 30, 2012 at 10:47 PM, Richard Graham
> <richardg at mellanox.com> wrote:
> > Actually, this is a working group that went out and spent quite a
> > bit of
> effort to collect user input. The issue is that it took a couple of
> years until someone had time to start an implementation, then took
> time for users to try it out and provide feedback, and then the
> broader forum started providing more input once text was actually written.
> >
> >
> >
> > It is true that this is important to HPC, however it is also true
> > that there are
> quite a few users outside of the HPC community that would like to use
> MPI, but can't because most/all MPI's currently terminate on process
> failure. We had users from the HPC community attend for a while,
> users with enterprise needs, and we even had input from an online
> gaming company. Even as recently as a couple of weeks ago this same
> point was brought up by some at Mellanox, when this was not even the
> topic of discussion. This is far from just a research activity of interest to a small number of users.
> >
> >
> >
> > Rich
> >
> >
> >
> > From: mpi3-ft-bounces at lists.mpi-forum.org
> > [mailto:mpi3-ft-bounces at lists.mpi-forum.org] On Behalf Of Ralph
> > Castain
> > Sent: Thursday, May 31, 2012 11:45 AM
> >
> >
> > To: MPI 3.0 Fault Tolerance and Dynamic Process Control working
> > Group
> > Subject: Re: [Mpi3-ft] Ticket 323 - status?
> >
> >
> >
> > Obviously, I can't speak for the folks who attended and voted "no",
> > either
> directly or by abstaining. However, I have talked to at least a few
> people, and can offer a point or two about the concerns.
> >
> > First, the last study I saw published on the subject of FT for MPI
> > showed a
> very low level of interest in FT within the MPI community. It based
> this on a usage analysis that showed something over 90% of clusters
> being too small to see large failure rates. On the clusters that were
> large enough (primarily at the national labs, who pretty clearly voted
> no), over 80% of the MPI jobs lasted less than 1 hour.
> >
> > So the size of the community that potentially benefits from FT is very small.
> In contrast, despite assurances it would be turned off unless
> specifically requested, it was clear from the proposals that FT would
> impact a significant fraction of the code, thus raising the potential
> for a substantial round of debugging and instability.
> >
> > For that majority who would see little-to-no benefit, this isn't an
> > attractive
> trade-off.
> >
> > Second, those who possibly could benefit tend to take a more
> > holistic view
> of FT. If you step back and look at the cluster as a system, then
> there are multiple ways of addressing the problems of failure during
> long runs. Yes, one way is to harden MPI to such events, but that is
> probably the hardest solution.
> >
> > One easier way, and the one being largely touted at the moment, is
> > to
> make checkpointing of an application be a relatively low-cost event so
> that it can be frequently done. This is being commercialized as we
> speak by the addition of SSDs to the usual parallel file system, thus
> making a checkpoint run at very fast speeds. In fact, "burst" buffers
> are allowing the checkpoint to dump very quickly, and then slowly
> drain to disk, rendering the checkpoint operation very low cost. Given
> that the commercial interests coincide with the HPC interests, this
> solution is likely to be available from cluster suppliers very soon at an attractive price.
> >
> > Combined with measures to make restart very fast as well, this looks
> > like an
> alternative that has no performance impact on the application at the
> MPI level, doesn't potentially destabilize the software, and may meet
> the majority of needs.
> >
> > I'm not touting this approach over any other, mind you - just trying
> > to point
> out that the research interests of the FT/MPI group needs to be
> considered in a separate light from the production interests of the
> community. What you may be experiencing (from my limited survey) is a
> reflection of that divergence.
> >
> > Ralph
> >
> >
> > On Wed, May 30, 2012 at 6:55 PM, George Bosilca
> > <bosilca at eecs.utk.edu>
> wrote:
> >
> > On May 31, 2012, at 08:44 , Martin Schulz wrote:
> >
> >
> >
> >
> > Several people who abstained had very similar concerns, but chose
> > the abstain vote since they know it meant no,
> >
> >
> >
> > Your interpretation is barely making a "simple majority" in the
> > forum, as
> highlighted by parallel discussions in the other email threads. But
> let's leave this discussion in its own thread.
> >
> >
> >
> > But you're right, both "no" and "abstain" votes should be addressed.
> > Bill
> made his point very clear, and to be honest he was the only one that
> raised a __valid__ point about the FT proposal. Personally, I am
> looking forward to fruitful discussions during our weekly phone-calls
> where the complaints raised during the voting will be brought forward
> in the way that the working group will have a real opportunity to
> address them as they deserve. In other terms we are all counting on
> you guys to enlighten us on the major issues with this proposal and the potential solutions you envision or promote.
> >
> >
> >
> > george.
> >
> >
> >
> > On May 31, 2012, at 08:44 , Martin Schulz wrote:
> >
> >
> >
> >
> > Hi George,
> >
> >
> >
> > One other no was Intel as far as I remember, but I don't remember the 5th.
> However, I would suggest not to focus on the no votes alone. Several
> people who abstained had very similar concerns, but chose the abstain
> vote since they know it meant no, but they agreed with the general
> necessity of FT for MPI. I remember, for example, Bill saying that for
> him abstain meant no, but that changes later on could change his mind.
> Based on this interpretation, the ticket definitely had more than 5 no votes.
> >
> >
> >
> > Martin
> >
> >
> >
> >
> >
> > On May 31, 2012, at 8:34 AM, Darius Buntinas wrote:
> >
> >
> >
> >
> >
> > Argonne was not convinced that we (FTWG) had the right solution, and
> > the
> large changes in the text mentioned previously did not instill
> confidence. So it was decided that Argonne would vote against the ticket.
> >
> > -d
> >
> > On May 30, 2012, at 6:24 PM, George Bosilca wrote:
> >
> >
> >
> > In total there were 5 no votes. I wonder who were the other two,
> > they
> might be willing to enlighten us on their reasons to vote against.
> >
> >
> >
> > george.
> >
> >
> >
> > On May 31, 2012, at 05:48 , Anthony Skjellum wrote:
> >
> >
> >
> > Three no votes were LLNL, Argonne, and Sandia. Since MPI is heavily
> driven by DOE, convincing these folks would be important.
> >
> >
> >
> > Tony Skjellum, tonyskj at yahoo.com or skjellum at gmail.com
> >
> > Cell 205-807-4968
> >
> >
> >
> > On May 31, 2012, at 5:10 AM, Richard Graham <richardg at mellanox.com>
> wrote:
> >
> >
> >
> > The main objection raised is that the text has still been having
> > large
> changes, and if not for the pressure of the 3.0 deadline, this would
> not have come up for a vote. I talked one-on-one with many that
> either voted against or abstained, and this was the major (not only) point raised.
> >
> >
> >
> > Rich
> >
> >
> >
> > -----Original Message-----
> >
> > From: mpi3-ft-bounces at lists.mpi-forum.org
> > [mailto:mpi3-ft-bounces at lists.mpi-forum.org] On Behalf Of Aurélien
> > Bouteiller
> >
> > Sent: Wednesday, May 30, 2012 10:05 PM
> >
> > To: MPI 3.0 Fault Tolerance and Dynamic Process Control working
> > Group
> >
> > Subject: Re: [Mpi3-ft] Ticket 323 - status?
> >
> >
> >
> > It seems we had very little, if any, technical opposition on the
> > content of
> the proposal itself, but mostly comments on the process. I think we
> need to understand more what are the oppositions. Do we have a list of
> who voted for and against and their rationale?
> >
> >
> >
> > Aurelien
> >
> >
> >
> >
> >
> > Le 30 mai 2012 à 08:52, Josh Hursey a écrit :
> >
> >
> >
> > That is unfortunate. A close vote (7 yes to 9 no/abstain). :/
> >
> >
> >
> > Thanks,
> >
> > Josh
> >
> >
> >
> > On Wed, May 30, 2012 at 8:38 AM, Thomas Herault
> >
> > <herault.thomas at gmail.com> wrote:
> >
> > Le 30 mai 2012 a 01:44, George Bosilca a écrit:
> >
> >
> >
> > The ticket has been voted down. Come back in 6 months, maybe 3.1.
> > The
> votes were 7 yes, 4 abstains and 5 no.
> >
> >
> >
> > Thomas
> >
> >
> >
> > Le 30 mai 2012 à 07:02, Josh Hursey a écrit :
> >
> >
> >
> > How did the vote go for the fault tolerance ticket 323?
> >
> >
> >
> > -- Josh
> >
> >
> >
> > --
> >
> > Joshua Hursey
> >
> > Postdoctoral Research Associate
> >
> > Oak Ridge National Laboratory
> >
> > http://users.nccs.gov/~jjhursey
> >
> > _______________________________________________
> >
> > mpi3-ft mailing list
> >
> > mpi3-ft at lists.mpi-forum.org
> >
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> >
> >
> > _______________________________________________
> >
> > mpi3-ft mailing list
> >
> > mpi3-ft at 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
> >
> >
> >
> > _______________________________________________
> >
> > mpi3-ft mailing list
> >
> > mpi3-ft at lists.mpi-forum.org
> >
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> > --
> >
> > * Dr. Aurélien Bouteiller
> >
> > * Researcher at Innovative Computing Laboratory
> >
> > * University of Tennessee
> >
> > * 1122 Volunteer Boulevard, suite 350
> >
> > * Knoxville, TN 37996
> >
> > * 865 974 9375
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > _______________________________________________
> >
> > mpi3-ft mailing list
> >
> > mpi3-ft at lists.mpi-forum.org
> >
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> > _______________________________________________
> >
> > mpi3-ft mailing list
> >
> > mpi3-ft at lists.mpi-forum.org
> >
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> >
> >
> > _______________________________________________
> >
> > mpi3-ft mailing list
> >
> > mpi3-ft at lists.mpi-forum.org
> >
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> > _______________________________________________
> > mpi3-ft mailing list
> > mpi3-ft at lists.mpi-forum.org
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> >
> __________________________________________________________
> ____________
> > __ Martin Schulz, schulzm at llnl.gov, http://people.llnl.gov/schulzm
> > CASC @ Lawrence Livermore National Laboratory, Livermore, USA
> >
> >
> >
> >
> >
> >
> >
> > _______________________________________________
> > mpi3-ft mailing list
> > mpi3-ft at lists.mpi-forum.org
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> >
> > _______________________________________________
> > mpi3-ft mailing list
> > mpi3-ft at lists.mpi-forum.org
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> >
> >
> >
> > _______________________________________________
> > mpi3-ft mailing list
> > mpi3-ft at lists.mpi-forum.org
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
> >
> > _______________________________________________
> > mpi3-ft mailing list
> > mpi3-ft at lists.mpi-forum.org
> > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
>
> --
> * Dr. Aurélien Bouteiller
> * Researcher at Innovative Computing Laboratory
> * University of Tennessee
> * 1122 Volunteer Boulevard, suite 309b
> * Knoxville, TN 37996
> * 865 974 9375
>
>
>
>
>
>
_______________________________________________
mpi3-ft mailing list
mpi3-ft at lists.mpi-forum.org
http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
More information about the mpiwg-ft
mailing list