On Apr 27, 2013, at 7:37 PM, Orion Poplawski <or...@cora.nwra.com> wrote:
> On 04/26/2013 08:53 PM, Ralph Castain wrote: >> >> On Apr 26, 2013, at 7:40 PM, Orion Poplawski <or...@cora.nwra.com> wrote: >> >>> On 04/25/2013 04:48 AM, Jeff Squyres (jsquyres) wrote: >>>> Orion -- >>>> >>>> I would just caution against using Open MPI with a thread-enabled >>>> libevent. In most cases, the performance impact wouldn't matter, but the >>>> whole point of MPI is to be high performance. The current 1.7 series does >>>> *not* use a thread-enabled libevent because it detracts from performance. >>>> Hence, using a thread-enabled libevent detracts from Open MPI's main >>>> purpose. >>>> >>>> If Open MPI is suddenly bundled with a thread-enabled libevent, >>>> *performance will go down* and users will be unhappy. We have learned >>>> painfully over the years that users expect good performance out of the box >>>> -- if they have get "bad" performance out of the box and have to do >>>> something special to enable "good" performance, they'll be annoyed and >>>> blame Open MPI. >>>> >>>> So I would request that you do *not* link Open MPI against a >>>> thread-enabled libevent until we are able integrate such functionality >>>> properly, and take measures to mitigate the performance implications >>>> (which likely won't be until at least the 1.9 series). >>> >>> So, the Fedora Packaging committee has taken that part to heart, but not >>> the bundling: >>> >>> * Bundling exception for libevent in openmpi - >>> https://fedorahosted.org/fpc/ticket/273 (spot, 16:16:52) >>> * ACTION: FPC is pretty universally against this bundling, spot will >>> try to make a libevent-nothread.so and update libevent. (spot, >>> 16:26:38) >> >> Just for clarification: who or what is "spot"? > > Tom "spot" Callaway - one the packaging committee members > >>> >>> So it looks like I will need to shortly be looking at how to link against >>> an external libevent. Any help with that would be greatly appreciated. >> >> As I said, I'll take a look at it, but can't commit to having it available >> any time soon. It isn't something I would suggest someone try who isn't >> fully versed in OMPI's code base. > > Yeah, I'm not looking forward to it. I get to at least wait until the > non-threaded version of libevent is available. I hate to see someone suffer, so I went ahead and added the external libevent connection this morning. Not trivial, but it seems to work. It is in our developer's trunk if you want to test it. As Jeff has said, we would prefer you not do this until the 1.9 series is released, and we won't be porting this change to the 1.7 series anyway. Just put it in so we can begin the investigation, and we always appreciate input and help in exploring the impacts! Ralph > > -- > Orion Poplawski > Technical Manager 303-415-9701 x222 > NWRA/CoRA Division FAX: 303-415-9702 > 3380 Mitchell Lane or...@cora.nwra.com > Boulder, CO 80301 http://www.cora.nwra.com > _______________________________________________ > devel mailing list > de...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/devel
