Bug#555714: mpich2: Please add support for blcr

Pavan Balaji balaji at mcs.anl.gov
Wed Nov 11 14:01:10 UTC 2009


On 11/11/2009 06:08 AM, Lucas Nussbaum wrote:
>> Since BLCR is now in the archive it would be nice if mpich2 was built 
>> using this library to provide checkpoint support. The attached patch 
>> enables this at configure time.

Ooh, I'd not recommend that for two reasons:

1. BLCR support is experimental in the mpich2-1.2.x series. It's really
meant for the mpich2-1.3.x series, preview releases of which should
start coming out in Jan 2010.

2. Checkpoint restart support is, in general, an expensive operation.
Even if the application doesn't checkpoint, the internal bookkeeping
required will hurt performance a little. So, even in the mpich2-1.3.x
series, I'm not sure how good an idea it is to enable it by default. You
might instead want to consider providing mpich2-blcr as a separate
package? Alternatively, we can consider protecting blcr based code by a
runtime option, as opposed to a compile-time option. This will still
hurt performance a little, but might be a lot more manageable. Let me
check with the developer who is adding this support and see what he thinks.

 -- Pavan

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





More information about the debian-science-maintainers mailing list