[Mpi3-rma] Target displacement sign issue

[Jeff Hammond]

I imagine that HPC-oriented operating systems do not provide virtual
addresses in the problematic range anyways.  Maybe VM randomization
causes this but that's not common in HPC.

In any case, MPI RMA requires MPI_Alloc_mem for portable performance.
MPI_Alloc_mem should be capable of returning only pointers in the
lower half of the address space, if by no other means than using mmap
intelligently.

I believe that this is only a problem on 32-bit systems where losing 2
of the 4 GB is a problem.  On a 64-bit system, losing half of the 18
exabytes in the address space is unlikely to be a problem, given that
an exabyte of DRAM would require something like a white dwarf to
power, meaning that it is likely that the OS can find a way to VM map
the physical memory present into the lower half of the address range.

Jeff

On Tue, Aug 28, 2012 at 12:31 PM, Jim Dinan <dinan at mcs.anl.gov> wrote:
> This would mean that the user can't use half of the address space for
> dynamic windows.  Not really ideal.
>
>  ~Jim.
>
>
> On 8/28/12 10:05 AM, Jeff Hammond wrote:
>>
>> Is it possible for implementations to prevent the user from attaching
>> memory to a dynamic window that would cause this overflow?  For
>> example, MPI_Win_attach could fail if the base+size indicated that
>> addresses could be used that were greater than the capacity of
>> MPI_Aint?  That seems like a way better option than breaking backwards
>> compatibility or attaching all sorts of caveats to the use of MPI_Aint
>> that a user is never going to remember.
>>
>> Jeff
>>
>> On Tue, Aug 28, 2012 at 10:00 AM, Jim Dinan <dinan at mcs.anl.gov> wrote:
>>>
>>> Sorry, this wasn't too clear.  A more detailed explanation of the problem
>>> is:
>>>
>>> Passing a dynamic window displacement to an RMA operation currently
>>> requires
>>> the user to put an address (which is an unsigned quantity) into an
>>> MPI_Aint
>>> (which is signed).  With the current spec, passing a negative
>>> displacement
>>> to an RMA operation is an error.  When interpreted as signed integers,
>>> dynamic window displacements (i.e. addresses) can be negative because the
>>> unsigned to signed conversion effectively overflows.
>>>
>>>   ~Jim.
>>>
>>>
>>> On 8/28/12 9:49 AM, Jeff Hammond wrote:
>>>>
>>>>
>>>> I do not understand this:
>>>>
>>>> "displacements associated with dynamic windows are addresses
>>>> (unsigned) and can range into the space of negative integers"
>>>>
>>>> How can they "range into the space of negative integers" if they are
>>>> unsigned?
>>>>
>>>> Jeff
>>>>
>>>> On Tue, Aug 28, 2012 at 9:46 AM, Jim Dinan <dinan at mcs.anl.gov> wrote:
>>>>>
>>>>>
>>>>> Hi All,
>>>>>
>>>>> Currently, target displacements in put/get/acc/etc are defined to be
>>>>> non-negative integers of type, MPI_Aint.  MPI_Aint is defined to be a
>>>>> signed
>>>>> integer.  However, displacements associated with dynamic windows are
>>>>> addresses (unsigned) and can range into the space of negative integers.
>>>>>
>>>>> It seems like the least impact change would be to add a special case
>>>>> that
>>>>> the displacement argument will be interpreted as unsigned for dynamic
>>>>> windows.  Changing the interface to take an 'unsigned MPI_Aint' would
>>>>> be
>>>>> the
>>>>> most clear (and preserve the existing and intended new semantics), but
>>>>> would
>>>>> break backward compatibility.
>>>>>
>>>>> Interested to hear your thoughts.
>>>>>
>>>>>    ~Jim.
>>>
>>>
>>> _______________________________________________
>>> mpi3-rma mailing list
>>> mpi3-rma at lists.mpi-forum.org
>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
>>
>>
>>
>>
> _______________________________________________
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond