[Mpi3-ft] Fault recovery of multiple communication libraries
I would like to understand how MPI can be notified of a failure. Consider another communication library (ARMCI/GASnet/...) identifying a failure through its own mechanisms. In the current model, how can it notify MPI to verify/reconfigure/recover from the error, for example by performing an MPI communication to the failed process? Conversely, can a communication library register to be notified of an error that MPI identifies and recovers from, so that that the library can take appropriate action? Sriram.K -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of [email protected] Sent: Thursday, February 12, 2009 11:17 AM To: [email protected] Subject: mpi3-ft Digest, Vol 13, Issue 4 Send mpi3-ft mailing list submissions to [email protected] To subscribe or unsubscribe via the World Wide Web, visit http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft or, via email, send a message with subject or body 'help' to [email protected] You can reach the person managing the list at [email protected] When replying, please edit your Subject line so it is more specific than "Re: Contents of mpi3-ft digest..." Today's Topics: 1. Re: Communicator Virtualization as a step forward (Josh Hursey) 2. Re: Communicator Virtualization as a step forward (Graham, Richard L.) 3. Re: Communicator Virtualization as a step forward (George Bosilca) 4. Re: Communicator Virtualization as a step forward (Nathan DeBardeleben) ---------------------------------------------------------------------- Message: 1 Date: Thu, 12 Feb 2009 12:23:49 -0500 From: Josh Hursey <[email protected]> Subject: Re: [Mpi3-ft] Communicator Virtualization as a step forward To: "MPI 3.0 Fault Tolerance and Dynamic Process Control working Group" <[email protected]> Message-ID: <[email protected]> Content-Type: text/plain; charset=US-ASCII; format=flowed; delsp=yes Yeah I was planning on using that document as a starting point. I wanted to look over it again and see if anything needs to change given some of the discussions that we have been having in the group. It may also need some additional language about MPI2 interfaces. It has been while since I have looked over this particular document. I would also like to add some more control for process startup, but we may decide to take that on as a secondary step. Does UTK still have the LaTeX for that document somewhere? Do you know if Graham would be interested in participating in this development? Cheers, Josh On Feb 12, 2009, at 10:43 AM, George Bosilca wrote:
Josh, the document that you talk about already exist. It was published
in ISC'04. Here is the link:http://www.netlib.org/utk/people/JackDongarra/PAPERS/isc2004-FT-M PI.pdf
george.
------------------------------ Message: 2 Date: Thu, 12 Feb 2009 14:01:15 -0500 From: "Graham, Richard L." <[email protected]> Subject: Re: [Mpi3-ft] Communicator Virtualization as a step forward To: [email protected] Message-ID: <[email protected]> Content-Type: text/plain; charset=UTF-8 Josh, Very early on in the process we got feedback from users that an ft-mpi like interface was of no interest to them. They would just as soon terminate the application and restart rather than use this sort of approach. Having said that, there is already previous demonstration that the ft-mpi approach is useful for some applications. If you look closely at the spec, the ft-mpi approach is a subset. of the current subset. I am working on pulling out the api's and expanding the explanations. The goal is to have this out before the next telecon in two weeks. Prototyping is under way, with ut, cray, and ornl committed to working on this. Right now supporting infrastructure is being developed. Your point on the mpi 2 interfaces is good. A couple of people had started to look at this when it looked like this might make it into the 2.2 version. The changes seemed to be more extensive than expected, so work stopped. This does need to be picked up on. Rich ------Original Message------ From: Josh Hursey To: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group ReplyTo: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group Sent: Feb 12, 2009 8:31 AM Subject: Re: [Mpi3-ft] Communicator Virtualization as a step forward It is a good point that local communicator reconstruction operations require a fundamental change in the way communicators are handled by MPI. With that in mind it would probably take as much effort (if not more) to implement a virtualized version on top of MPI. So maybe it will not help as much as I had originally thought. Outside of the paper, do we have the interface and semantics of these operations described anywhere? I think that would help in trying to keep pace with the use cases. The spirit of the suggestion was as a way to separate what (I think) we can agree on as a first step (FT-MPI-like model) from the communicator reconstruction, which I see as a secondary step. If we stop to write up what the FT-MPI-like model should look like in the standard, then I think we can push forward on other fronts (prototyping of step 1, standardization of step 1, application implementations using step 1) while still trying to figure out how communication reconstruction should be expressed in the standard such that it is usable in target applications. So my motion is that the group explicitly focus effort on writing a document describing the FT-MPI-like model we consider as a foundation. Do so in the MPI standard language, and present it to the MPI Forum for a straw vote in the next couple of meetings. From this document we can continue evolving it to support more advanced features, like communicator reconstruction. I am willing to put effort into making such a document. However, I would like explicit support from the working group in pursing such an effort, and the help of anyone interested in helping write-up/define this specification. So what do people think taking this first step? -- Josh On Feb 11, 2009, at 5:57 PM, Greg Bronevetsky wrote:
I don't understand what you mean by "We can continue to pursue communicator reconstruction interfaces though a virtualization later above MPI." To me it seems that such interfaces will effectively need
to implement communicators on top of MPI in order be operational, which will take about as much effort as implementing them inside MPI. In particular, I don't see a way to recreate a communicator using the MPI interface without making collective calls. However, we're defining
MPI_Rejoin (or whatever its called) to be a local operation. This means that we cannot use the MPI communicators interface and must instead implement our own communicators.
The bottom line is that it does make sense to start implementing support for the FT-MPI model and evolve that to a more elaborate model. However, I don't think that working on the rest above MPI will save us any effort or time.
Greg Bronevetsky Post-Doctoral Researcher 1028 Building 451 Lawrence Livermore National Lab (925) 424-5756 [email protected]
At 01:17 PM 2/11/2009, Josh Hursey wrote:
In our meeting yesterday, I was sitting in the back trying to take in
the complexity of communicator recreation. It seems that much of the confusion at the moment is that we (at least I) are still not exactly
sure how the interface should be defined and implemented.
I think of the process fault tolerance specification as a series of steps that can be individually specified building upon each step while working towards a specific goal set. From this I was asking myself, is there any foundational concepts that we can define now so that folks can start implementation.
That being said I suggest that we consider FT-MPI's model of all communicators except the base 3 (COMM_WORLD, COMM_SELF, COMM_NULL) are destroyed on a failure as the starting point for implementation. This would get us started. We can continue to pursue communicator reconstruction interfaces though a virtualization later above MPI. We
can use this layer to experiment with the communicator recreation mechanisms in conjunction with applications while pursing the first step implementation. Once we start to agree on the interface for communicator reconstruction, then we can start to push it into the MPI standard/library for a better standard/implementation.
The communicator virtualization library is a staging area for these interface ideas that we seem to be struggling with. The virtualization
------Original Message Truncated------ ------------------------------ Message: 3 Date: Thu, 12 Feb 2009 14:16:09 -0500 From: George Bosilca <[email protected]> Subject: Re: [Mpi3-ft] Communicator Virtualization as a step forward To: "MPI 3.0 Fault Tolerance and Dynamic Process Control working Group" <[email protected]> Message-ID: <[email protected]> Content-Type: text/plain; charset=US-ASCII; format=flowed; delsp=yes I don't necessarily agree with the statement that FT-MPI is a subset of the current spec. As the current spec can be implemented on top of FT-MPI (with help from the PMPI interface), this tend to prove the opposite. However, I agree there are several features in the current spec that were not covered by the FT-MPI spec, but these features can be implemented on top of FT-MPI. As far as I understood, this is what Josh proposed, as this will give a quick start (i.e. FT-MPI implementation is already available). george. On Feb 12, 2009, at 14:01 , Graham, Richard L. wrote:
Josh, Very early on in the process we got feedback from users that an ft- mpi like interface was of no interest to them. They would just as soon terminate the application and restart rather than use this sort of approach. Having said that, there is already previous demonstration that the ft-mpi approach is useful for some applications. If you look closely at the spec, the ft-mpi approach is
a subset. of the current subset. I am working on pulling out the api's and expanding the explanations.
The goal is to have this out before the next telecon in two weeks. Prototyping is under way, with ut, cray, and ornl committed to working on this. Right now supporting infrastructure is being developed. Your point on the mpi 2 interfaces is good. A couple of people had started to look at this when it looked like this might make it into the 2.2 version. The changes seemed to be more extensive than expected, so work stopped. This does need to be picked up on.
Rich ------Original Message------ From: Josh Hursey To: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group ReplyTo: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group Sent: Feb 12, 2009 8:31 AM Subject: Re: [Mpi3-ft] Communicator Virtualization as a step forward
It is a good point that local communicator reconstruction operations require a fundamental change in the way communicators are handled by MPI. With that in mind it would probably take as much effort (if not more) to implement a virtualized version on top of MPI. So maybe it will not help as much as I had originally thought. Outside of the paper, do we have the interface and semantics of these operations described anywhere? I think that would help in trying to keep pace with the use cases.
The spirit of the suggestion was as a way to separate what (I think) we can agree on as a first step (FT-MPI-like model) from the communicator reconstruction, which I see as a secondary step. If we stop to write up what the FT-MPI-like model should look like in the standard, then I think we can push forward on other fronts (prototyping of step 1, standardization of step 1, application implementations using step 1) while still trying to figure out how communication reconstruction should be expressed in the standard such that it is usable in target applications.
So my motion is that the group explicitly focus effort on writing a document describing the FT-MPI-like model we consider as a foundation.
Do so in the MPI standard language, and present it to the MPI Forum for a straw vote in the next couple of meetings. From this document we
can continue evolving it to support more advanced features, like communicator reconstruction.
I am willing to put effort into making such a document. However, I would like explicit support from the working group in pursing such an effort, and the help of anyone interested in helping write-up/define this specification.
So what do people think taking this first step?
-- Josh
On Feb 11, 2009, at 5:57 PM, Greg Bronevetsky wrote:
I don't understand what you mean by "We can continue to pursue communicator reconstruction interfaces though a virtualization later above MPI." To me it seems that such interfaces will effectively need to implement communicators on top of MPI in order be operational, which will take about as much effort as implementing them inside MPI. In particular, I don't see a way to recreate a communicator using the MPI interface without making collective calls.
However, we're defining MPI_Rejoin (or whatever its called) to be a local operation. This means that we cannot use the MPI communicators interface and must instead implement our own communicators.
The bottom line is that it does make sense to start implementing support for the FT-MPI model and evolve that to a more elaborate model. However, I don't think that working on the rest above MPI will
save us any effort or time.
Greg Bronevetsky Post-Doctoral Researcher 1028 Building 451 Lawrence Livermore National Lab (925) 424-5756 [email protected]
At 01:17 PM 2/11/2009, Josh Hursey wrote:
In our meeting yesterday, I was sitting in the back trying to take in the complexity of communicator recreation. It seems that much of the confusion at the moment is that we (at least I) are still not exactly sure how the interface should be defined and implemented.
I think of the process fault tolerance specification as a series of steps that can be individually specified building upon each step while working towards a specific goal set. From this I was asking myself, is there any foundational concepts that we can define now so
that folks can start implementation.
That being said I suggest that we consider FT-MPI's model of all communicators except the base 3 (COMM_WORLD, COMM_SELF, COMM_NULL) are destroyed on a failure as the starting point for implementation. This would get us started. We can continue to pursue communicator reconstruction interfaces though a virtualization later above MPI. We can use this layer to experiment with the communicator recreation mechanisms in conjunction with applications while pursing the first step implementation. Once we start to agree on the interface for communicator reconstruction, then we can start to push it into the MPI standard/library for a better standard/implementation.
The communicator virtualization library is a staging area for these interface ideas that we seem to be struggling with. The virtualization
------Original Message Truncated------
_______________________________________________ mpi3-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
------------------------------ Message: 4 Date: Thu, 12 Feb 2009 12:16:23 -0700 From: Nathan DeBardeleben <[email protected]> Subject: Re: [Mpi3-ft] Communicator Virtualization as a step forward To: "MPI 3.0 Fault Tolerance and Dynamic Process Control working Group" <[email protected]> Message-ID: <[email protected]> Content-Type: text/plain; charset=ISO-8859-1; format=flowed I really worry about taking the advise of users saying they would rather terminate and restart an application than having some assistance to help them ride through a problem. If they are worried about programming language/model changes, I would encourage them to open their eyes. Major programming model changes are predicted for > petascale computers and even petascale computers are having a hard time with classical MPI programming. I think we're more likely to see MPI as an underpinning of next-gen models. These users polled might not be extreme-scale users, however. Working at a laboratory positioning itself for exascale, we are intimately aware of the fact that "oh just rerun it" is a worthless conclusion. I wish I had more time to assist in this matter but our laboratory has cracked down on participation in things that are not directly associated with charge codes so it's a bit hard for me to spend any sizable amount of time. Please though, consider the user base when they say things like that. I'm sure Rich is well aware of these similar concerns. While MPI fault tolerance might not be important to the users running 1000 node systems, those of us approaching system mean time to interrupt under an hour are quite on the opposite side of that spectrum. Are the small-system users pushing for FT to not be inside of MPI? This is why I was so in favor of some sort of componentized MPI where users could exclude FT if they weren't worried about reliability (and thereby gain performance) but those of us who were in more dangerous reliability regimes could take the performance penalty and compile in / load in / configure in / whatever FT. -- Nathan --------------------------------------------------------------------- Nathan DeBardeleben, Ph.D. Los Alamos National Laboratory High Performance Computing Systems Integration (HPC-5) phone: 505-667-3428 email: [email protected] --------------------------------------------------------------------- Graham, Richard L. wrote:
Josh, Very early on in the process we got feedback from users that an ft-mpi like interface was of no interest to them. They would just as soon terminate the application and restart rather than use this sort of approach. Having said that, there is already previous demonstration that the ft-mpi approach is useful for some applications. If you look closely at the spec, the ft-mpi approach is a subset. of the current subset. I am working on pulling out the api's and expanding the explanations. The goal is to have this out before the next telecon in two weeks. Prototyping is under way, with ut, cray, and ornl committed to working on this. Right now supporting infrastructure is being developed. Your point on the mpi 2 interfaces is good. A couple of people had started to look at this when it looked like this might make it into the 2.2 version. The changes seemed to be more extensive than expected, so work stopped. This does need to be picked up on.
Rich ------Original Message------ From: Josh Hursey To: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group ReplyTo: MPI 3.0 Fault Tolerance and Dynamic Process Control working Group Sent: Feb 12, 2009 8:31 AM Subject: Re: [Mpi3-ft] Communicator Virtualization as a step forward
It is a good point that local communicator reconstruction operations require a fundamental change in the way communicators are handled by MPI. With that in mind it would probably take as much effort (if not more) to implement a virtualized version on top of MPI. So maybe it will not help as much as I had originally thought. Outside of the paper, do we have the interface and semantics of these operations described anywhere? I think that would help in trying to keep pace with the use cases.
The spirit of the suggestion was as a way to separate what (I think) we can agree on as a first step (FT-MPI-like model) from the communicator reconstruction, which I see as a secondary step. If we stop to write up what the FT-MPI-like model should look like in the standard, then I think we can push forward on other fronts (prototyping of step 1, standardization of step 1, application implementations using step 1) while still trying to figure out how communication reconstruction should be expressed in the standard such that it is usable in target applications.
So my motion is that the group explicitly focus effort on writing a document describing the FT-MPI-like model we consider as a foundation.
Do so in the MPI standard language, and present it to the MPI Forum for a straw vote in the next couple of meetings. From this document we
can continue evolving it to support more advanced features, like communicator reconstruction.
I am willing to put effort into making such a document. However, I would like explicit support from the working group in pursing such an effort, and the help of anyone interested in helping write-up/define this specification.
So what do people think taking this first step?
-- Josh
On Feb 11, 2009, at 5:57 PM, Greg Bronevetsky wrote:
I don't understand what you mean by "We can continue to pursue communicator reconstruction interfaces though a virtualization later above MPI." To me it seems that such interfaces will effectively need to implement communicators on top of MPI in order be operational, which will take about as much effort as implementing them inside MPI. In particular, I don't see a way to recreate a communicator using the MPI interface without making collective calls.
However, we're defining MPI_Rejoin (or whatever its called) to be a local operation. This means that we cannot use the MPI communicators interface and must instead implement our own communicators.
The bottom line is that it does make sense to start implementing support for the FT-MPI model and evolve that to a more elaborate model. However, I don't think that working on the rest above MPI will
save us any effort or time.
Greg Bronevetsky Post-Doctoral Researcher 1028 Building 451 Lawrence Livermore National Lab (925) 424-5756 [email protected]
At 01:17 PM 2/11/2009, Josh Hursey wrote:
In our meeting yesterday, I was sitting in the back trying to take in the complexity of communicator recreation. It seems that much of the confusion at the moment is that we (at least I) are still not exactly sure how the interface should be defined and implemented.
I think of the process fault tolerance specification as a series of steps that can be individually specified building upon each step while working towards a specific goal set. From this I was asking myself, is there any foundational concepts that we can define now so
that folks can start implementation.
That being said I suggest that we consider FT-MPI's model of all communicators except the base 3 (COMM_WORLD, COMM_SELF, COMM_NULL) are destroyed on a failure as the starting point for implementation.
This would get us started. We can continue to pursue communicator reconstruction interfaces though a virtualization later above MPI. We can use this layer to experiment with the communicator recreation
mechanisms in conjunction with applications while pursing the first step implementation. Once we start to agree on the interface for communicator reconstruction, then we can start to push it into the MPI standard/library for a better standard/implementation.
The communicator virtualization library is a staging area for these interface ideas that we seem to be struggling with. The virtualization
------Original Message Truncated------
_______________________________________________ mpi3-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
------------------------------ _______________________________________________ mpi3-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft End of mpi3-ft Digest, Vol 13, Issue 4 **************************************
participants (1)
-
Krishnamoorthy, Sriram