faq/tcp: add an entry related to 10GbE networks and faster

[ggouaillardet]

• set send and receive buffers size to zero
• try multiple links

Signed-off-by: Gilles Gouaillardet gilles@rist.or.jp

[ggouaillardet]

@bwbarrett @bosilca can you please double check this ?

this is related to https://www.mail-archive.com/users@lists.open-mpi.org/msg31283.html

if there are some recommended settings for AWS, they can be added and/or linked too

[bgoglin]

I don't understand "Open MPI < v2.0.3 or > v2.1.2".

[ggouaillardet]

oops, that is a pretty dumb typo i will fix right away

[bosilca]

I would rather be more specific, and not tie this capability to a specific version of OMPI, but instead talk about the real reason why this settings are not necessary anymore. We should say that in modern version of the Linux kernel, setting the TCP send and receive buffers is a unnecessary operation that can lead to inefficiencies. We suggest users to check their TCP buffer sizes using ompi_info --param btl tcp -l 4 and make sure that btl_tcp_sndbuf and btl_tcp_rcvbuf are set to 0. Starting from Open MPI 2.0.3 this are the default settings.

[bwbarrett]

I agree with George. Also, the btl_num_links option doesn't work correctly. We need to sort that out before we suggest it in a FAQ.

[ggouaillardet]

i updated this PR, please double check the wording

[gpaulsen]

@ggouaillardet, @bwbarrett @bosilca: Is this still relevant? Should we pull it in, or are we still waiting on some fixes?

[bwbarrett]

We need to test, because I think multilink works in 4.0. Either way, I think adding the multilinks answer is a really bad look for Open MPI.

[bosilca]

We need closure on this topic. I just added this topic to the developer's meeting agenda.

[bwbarrett]

Closing in favor of #111