[mpich-discuss] MPI_Alloc_mem: optimized memory usage

Dave Goodell goodell at mcs.anl.gov
Tue Nov 11 12:31:19 CST 2008


On Nov 11, 2008, at 11:14 AM, Dorian Krause wrote:

> Rajeev Thakur wrote:
>> Yes, it is useful for MPI-2 one-sided communication, but perhaps  
>> less useful
>> for send-recv.
>>
>> Rajeev
>>
>
> What would be the advantage on commodity hardware? Can we do better  
> than malloc if the Interconnect doesn't support something like RDMA?
>
> Thanks,
> Dorian

You could use mmap'ed or SysV shared memory, for example.  This could  
reduce the number of copies in some cases.  The exact benefit would  
depend on the particular scenario that you are trying to optimize.

-Dave


>>> -----Original Message-----
>>> From: mpich-discuss-bounces at mcs.anl.gov [mailto:mpich-discuss- 
>>> bounces at mcs.anl.gov] On Behalf Of James S Perrin
>>> Sent: Monday, November 10, 2008 5:24 AM
>>> To: mpich-discuss at mcs.anl.gov
>>> Subject: Re: [mpich-discuss] MPI_Alloc_mem: optimized memory usage
>>>
>>> Thanks for the info Rajeev. Would this be something worth looking  
>>> at in future releases in the nemesis channel as we run more  
>>> systems with more and more cores per node?
>>>
>>> Regards
>>> James
>>>
>>> Rajeev Thakur wrote:
>>>
>>>> In the sshm channel (scalable shared memory), which is now
>>> deprecated,
>>>
>>>> Alloc_mem allocated shared memory for use in one-sided
>>> communication. I
>>>
>>>> believe in other cases at present, Alloc_mem just calls malloc.
>>>>
>>>> Rajeev
>>>>
>>>>> -----Original Message-----
>>>>> From: mpich-discuss-bounces at mcs.anl.gov [mailto:mpich-discuss- 
>>>>> bounces at mcs.anl.gov] On Behalf Of
>>> James S Perrin
>>>
>>>>> Sent: Friday, November 07, 2008 11:55 AM
>>>>> To: mpich
>>>>> Subject: [mpich-discuss] MPI_Alloc_mem: optimized memory usage
>>>>>
>>>>> Hi,
>>>>> 	Are MPICH2 calls optimized for SMP/multi-core architectures if  
>>>>> I allocate buffers using MPI_Alloc_mem()? I was reading the in  
>>>>> the MPI2.1 spec that MPI_Alloc_mem can be implemented to take  
>>>>> advantage of shared memory pools when available meaning I can  
>>>>> keep my code
>>> portable and
>>>>> clean. Presumably this would only be available in ch3:ssm
>>> and nemesis
>>>>> channels.
>>>>>
>>>>> Regards
>>>>> James




More information about the mpich-discuss mailing list