yoda: avoid useless memcpy in mca_spml_yoda_get when the btl is get capa...

[ggouaillardet]

...ble

[ggouaillardet]

@miked-mellanox @hjelmn could you please review this ?

by setting getreq->p_dst to NULL, we avoid a useless memcpy (src==dst !)
in mca_spml_yoda_get_completion.
oshrun -np 2 --mca btl vader,self examples/oshmem_max_reduction
can be used to test with knem loaded.

[mellanox-github]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
http://bgate.mellanox.com/jenkins/job/gh-ompi-master-pr/41/
Test PASSed.

[mike-dubman]

@alex-mikheev - please comment.

[ggouaillardet]

@alex-mikheev at first glance that does make sense.

now when i run the oshmem_max_reduction test, the destination is not a symmetric object (it is a local variable in the stack) and :

1. registration is set to NULL since l_mkey is NULL
2. mca_spml_yoda_get invokes vader_prepare_dst, that will KNEM_CMD_CREATE_REGION for the local variable (in the case of vader with knem, this is useless, in a more general case, this might just be suboptimal, @hjelmn could you please comment ?)
3. so your proposed fix will not set getreq->p_dst to NULL so memcpy will be invoked with src==dst

[alex-mikheev]

Is there a way to know that btl can do get() to non registered memory or put() from non registered memory without requiring bounce buffers ?

[hjelmn]

@ggouaillardet The new btl interface will resolve issue #2. If you register a region for local write only vader will not create a knem cookie for the region (this is not implemented in the code yet-- will probably be implemented this weekend).

@alex-mikheev There is no way to tell ATM if a region can be used for BTL rdma without a registration key. I will make sure this can be determined in the new interface.

[jsquyres]

So this PR is awaiting #278.

[jsquyres]

@alex-mikheev @ggouaillardet @miked-mellanox Is this PR still relevant? If so, is it ready to go?

[hppritcha]

Can one of the admins verify this patch?

[hppritcha]

Where are we with this PR now? Given this is more than 6 months old seems to be time to decide whether to close or merge.

[lanl-ompi]

Refer to this link for build results (access rights to CI server needed):
http://jenkins.open-mpi.org/job/ompi_master_pr_distcheck/8/
Test PASSed.

[lanl-ompi]

Refer to this link for build results (access rights to CI server needed):
http://jenkins.open-mpi.org/job/ompi_master_pr_cle5.2up02/97/
Test PASSed.

[jsquyres]

Last activity on this PR was in 2014. I'm guessing it isn't relevant any more.

Feel free to re-open if that's and incorrect assumption.