[Mpi3-ft] Problem with reusing rendezvous memory buffers

Wed Jul 31 12:28:08 CDT 2013

George,

I'm not talking about RMA here.  I'm just talking about send/recv.  I'm 
also not talking about simultaneous receives posted on the same buffer. 
  They are separated by failure notifications, COMM_AGREE, whatever else 
you need.  I think Sayantan got the point Wesley mentioned and his 
solution is correct (assuming enough support from the network to do so). 
  Though I'm not convinced it's not adding overhead.  Sayantan should 
comment on this.

With respect to the K Computer, I don't have a link.  This was my 
understanding from what the Fujitsu folks mentioned (and was the reason 
why they didn't want to release their low-level API; since the hardware 
has no protection, anyone could write to anyone's memory).  And I'm not 
trying to prove that K computer is not good enough in any way.  That was 
just an example.  My point was only that you should consider the case 
that not all networks would have such protection capabilities.

  -- Pavan

On 07/31/2013 11:50 AM, George Bosilca wrote:
> On Tue, Jul 30, 2013 at 10:47 PM, Pavan Balaji <balaji at mcs.anl.gov> wrote:
>>
>> Hmm.  Yes, you are right.  Generating different per-process rkeys is an
>> option on IB.  Though that's obviously less scalable than a single rkey and
>> hurts performance too because a new rkey has to be generated for each
>> process.  Even more of a reason for FT to have a requested/provided option
>> like threads.
>
> I fail to understand the scenario allowing you to reach such a
> conclusion. You want to have an MPI_RECV with a buffer where multiple
> senders can do RMA operations. The only way this can be done in the
> context of the MPI standard is if each of the receives on this
> particular buffer are using non-contiguous datatypes. Thus, unlike
> what you suggest in your answer above, this is not hurting performance
> as you are already in a niche mode (I'm not even talking about the
> fact that usually non-contiguous datatypes conflicts with RMA
> operations). Moreover, you suppose that the detection of a dead
> process and the re-posting of the receive buffer can happen faster
> than an RMA message cross the network. The only potential case where
> such a scenario can happen is when multiple paths between the source
> and the destination exist, and the failure detection happen on one
> path while the RMA message took another one. This is highly improbable
> in most cases.
>
> There are too many ifs in this scenario to make it plausible. Even if
> we suppose that all those ifs will be true, as Rich said, this is an
> issue of [quality of] implementation not MPI standard. A high quality
> MPI implementation will delay reporting the process failure error on
> that particular MPI_RECV until all possible RMA from the dead process
> were either discarded by the network, or written to the memory.
>
>>
>> However, please also think about this problem for other networks that might
>> not have such hardware protection capabilities (K Computer comes to mind).
>
> K computer ? My understanding is that there are such capabilities in
> the TOFU network. I might be wrong thou, in which case I would
> definitively appreciate if you can you pinpoint me to a
> link/documentation that proves your point?
>
>> Maybe they cannot provide MPI-specified FT, and that would be fine.
>
> Not really, FT can be supported without overhead for the normal
> execution even for the types of netwrok you mention. The solution I
> presented above, uses the timeouts of the network layer to ensure no
> delivery can occur after the error reporting, by delaying the error
> reporting until all timeout occurred. Trivial to implement, and
> without impact on the normal execution path.
>
>    Thanks,
>      George.
>
>
>>
>>   -- Pavan
>>
>>
>> On 07/30/2013 02:59 PM, Sur, Sayantan wrote:
>>>
>>> Hi Wesley,
>>>
>>> Looks like your attachment didn’t make it through. Using IB, one can
>>> generate rkeys for each sender and just invalidate the key for the
>>> observed failed process. HW can just drop the “slow” message when it
>>> arrives. I’m assuming that generating keys should be fast in the future
>>> given that recently announced HW/firmware has support for on-demand
>>> registration. In any case, it is not a restriction of IB per se.
>>>
>>> Thanks,
>>>
>>> Sayantan
>>>
>>> *From:*mpi3-ft-bounces at lists.mpi-forum.org
>>> [mailto:mpi3-ft-bounces at lists.mpi-forum.org] *On Behalf Of *Wesley Bland
>>> *Sent:* Tuesday, July 30, 2013 11:04 AM
>>> *To:* MPI3-FT Working Group
>>> *Subject:* [Mpi3-ft] Problem with reusing rendezvous memory buffers
>>>
>>>
>>> Pavan pointed out a problem to me yesterday related to memory buffers
>>> used with rendezvous protocols. If a process passes a piece of memory to
>>> the library in an MPI_RECV and the library gives that memory to the
>>> hardware, where it is pinned, we can get into trouble if one of the
>>> processes that could write into that memory fails. The problem comes
>>> from a process sending a slow message and then dying. It is possible
>>> that the other processes could detect and handle the failure before the
>>> slow message arrives. Then when the message does arrive, it could
>>> corrupt the memory without the application having a way to handle this.
>>> My whiteboard example is attached as an image.
>>>
>>> We can't just unmap memory from the NIC when a failure occurs because
>>> that memory is still being used by another process's message. Some
>>> hardware supports unmapping memory for specific senders which would
>>> solve this issue, but some don't, such as InfiniBand, where the memory
>>> region just has a key and unmapping it removes it for all senders.
>>>
>>> This problem doesn't have a good solution (that I've come up with), but
>>> I did come up with a solution. We would need to introduce another error
>>> code (something like MPI_ERR_BUFFER_UNUSABLE) that would be able to tell
>>> the application that the buffer that the library was using is no longer
>>> usable because it might be corrupted. For some hardware, this wouldn't
>>> have to be returned, but for hardware where this isn't possible, the
>>> library could pass this error to the implementation to say that I need a
>>> new buffer in order to complete this operation. On the sender side, the
>>> operation would probably complete successfully since to it, the memory
>>> was still available. That means that there will be some rollback
>>> necessary, but that's up to the application to figure out.
>>>
>>> I know this is an expensive and painful solution, but this is all I've
>>> come up with so far. Thoughts from the group?
>>>
>>> Thanks,
>>>
>>> Wesley
>>>
>>>
>>>
>>> _______________________________________________
>>> mpi3-ft mailing list
>>> mpi3-ft at lists.mpi-forum.org
>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft
>>>
>>
>> --
>> Pavan Balaji
>> http://www.mcs.anl.gov/~balaji
>>
>> _______________________________________________
>> 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
>

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji