Bug#555714: mpich2: Please add support for blcr

[Pavan Balaji]

On 10/24/2010 11:55 AM, Pavan Balaji wrote:
> On 10/24/2010 11:42 AM, Lucas Nussbaum wrote:
>> On 23/10/10 at 12:51 -0500, Pavan Balaji wrote:
>>> On 10/23/2010 12:40 PM, Lucas Nussbaum wrote:
>>>> On 11/11/09 at 10:15 -0600, Pavan Balaji wrote:
>>>>> We had a short discussion on this within the MPICH group. We believe
>>>>> that we will be able to modify the code such that enabling it at
>>>>> configure time should not add any additional performance overhead
>>>>> for applications that don't need this support. So starting
>>>>> mpich2-1.3, you should be able to configure it in by default.
>>>>
>>>> Hi Pavan,
>>>>
>>>> What's the conclusion on this? Should I enable checkpointing by default?
>>>> (Context: http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=555714)
>>>
>>> There will be some overhead in enabling this, but I think that
>>> should be minimal. I'm checking with our developer who worked on
>>> this. I'll get back to you.
>>
>> Apparently, if you enable it at build time, it then fails at runtime if
>> blcr support is not present in the kernel. So I've left it disabled for
>> now.
>
> Yikes! That's a bug. We'll fix it.

This has been fixed in the 1.3.x branch and trunk. Here are the relevant 
patches:

https://trac.mcs.anl.gov/projects/mpich2/changeset/7437
https://trac.mcs.anl.gov/projects/mpich2/changeset/7438
https://trac.mcs.anl.gov/projects/mpich2/changeset/7439

We also tested the performance penalty of enabling it by default, and 
looks like the penalty is very small (within the noise). So you should 
be able to enable it by default.

  -- Pavan

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