[ofa-general] Re: [ewg] RFC: Do we wish to take MPI out of OFED?

[Sven-Thorsten Dietrich]

Hi,

I have just read over the lengthy discussion on the subject, and we have
had some internal discussion on this topic since the OFA meeting earlier
in the year.

Without going into the many individual points:

A Linux system in general has many many interdependencies, which are
overseen by individual package and subsystem maintainers.

They generally seem to manage just fine to keep things working
seamlessly.

When this is not the case, the issues are usually addressed based on
direct input from customers, distro, or dependent system maintainers.

Things become confusing when there are multiple POCs for a package, as
with MPI in OFED.

In regards to integrated QA, it should suffice to specify a version for
MPI that have been tested with OFED/RDMA, and this version is something
that OFED vendors can sort out, and communicate to distros.

The specific source, however does not need to be re-packaged.

Not packaging MPI in OFED does not eliminate the ability of the OFED
maintainers and OFED users to interact with MPI maintainers to report
issues. This notably, happens even with QA'd code, as its generally
impossible to anticipate all issues at this level of complexity

What is eliminated, is having to re-release the entire OFED package
because of changes or updates in MPI.

In summary, decoupling as much as possible, while referencing specific
versions in QA testing, is fundamental for any integrated system, and I
see no need for OFED to odd-man-out on this issue.

At Novell we are therefore in favor of decoupling MPI from OFED.

Regards,

Sven