[openib-general] Announcing the OpenFabrics Enterprise Distribution

[Shawn Hansen (shahanse)]

All,

The OpenFabrics Enterprise Working Group is pleased to announce the
creation of the OpenFabrics Enterprise Distribution (OFED).  This was
approved today by the Board of Directors.

OFED is a distribution of InfiniBand software that includes, or is a
superset of, the OpenFabrics 1.0 release, and adds other additional
software outside of the scope of the OpenFabrics release, such as MPI.
This development work will happen completely in the open in the
Enterprise Working Group.  To join this mailing list, please subscribe
at: http://openib.org/mailman/listinfo/openfabrics-ewg

The major reasons for this distribution are:

1) Need to unify vendors' snapshots of OpenFabrics release for
interoperability;
2) Need to package software outside the scope of the main OpenFabrics
release;
3) Need to coordinate vendors' bug fixes and prevent divergence.

For frequently asked questions, a FAQ is included below.  Please let us
know if you have any additional questions.  

Regards,

Shawn Hansen
Chair, Enterprise Working Group
Cisco Systems
shawn.hansen at cisco.com

--------------------------
Frequently Asked Questions
--------------------------

Q: What is the Enterprise Working Group?

The EWG is a group of hardware vendors that will sell products based on
OpenFabrics.  The purpose of this group is to coordinate how to provide
a single commercially supportable distribution of OpenFabrics software
to their customers that guarantees cross-vendor interoperability.

Q: Why is OFED required?

- Enterprise customers will have solution-level requirements that are
outside the scope of the 1.0 release, such as the distribution of MPI
stacks, support for pre-2.6.16 kernels, etc.  The goal of OFED is to
address this need.  Without OFED, each InfiniBand vendor would create
their own distribution of OpenFabrics to accomplish this goal, and may
not be interoperable.

Q: Does OFED compete with the OpenFabrics release?

- No, there is only one OpenFabrics release.  OFED is a distribution
that includes the OpenFabrics 1.0 release. The OpenFabrics 1.0 release
and OFED share the same user-level code (libraries, management
utilities, etc.)  The code for both is taken from the 1.0 branch.

Q: Is OFED development happening in the open?

- Yes, OFED uses the OpenFabrics bugzilla for bug reporting, and all
discussions can be viewed on the Enterprise Working Group mailing list.
All OFED development is done on the 1.0 branch under the ibed directory.
Anyone can access release candidates, test them, observe bugs and
discussions, report bugs, and comment.

Q: How does OFED differ from the OpenFabrics release?

- The OpenFabrics release contains only user-level code, while the OFED
distribution also adds InfiniBand kernel modules that are under
OpenFabrics development, including modules that are not part of the
kernel (like iSER, RDS, and SDP).

- OFED will include two MPI packages that are not part of Open Fabrics:
OSU MPI and Open MPI.

- OFED is packaged for end-user installation.

- OFED supports distribution with older kernels (e.g. Redhat EL4 up2)

Q: What is the software release process for OFED and how does it relate
to the OpenFabrics release?

The release build is done using the following method:

1.  Any module that is already in the kernel will be taken from the git
tree that is targeted for next kernel release

2.  Kernel modules that are not in Linux kernel will be taken from
openFabrics SVN trunk or in extraordinary cases, from SVN contrib. 

3.  All user space code is taken from the 1.0 branch. OFED group will
make sure the right patches from the trunk are updated to the branch.

4.  MPI:
Open MPI - Provided by OpenMPI developers.
MVAPICH - Based on OSU release.
Both tarballs are placed in OpenFabrics web site.

5.  OFED build & install scripts: all relevant scripts are placed under
a specific directory for OFED release under the 1.0 branch. 

6.  Back port patches: patches directory will be also under the OFED
directory in the 1.0 branch. 

The release process:

The release coordinator will build the release candidate (OFED-rcX) and
publish it on OpenFabrics (approximately every 2 weeks).

Each OFED vendor is responsible to test the components under his
ownership. Bugs are reported through bugzilla and fixes are provided to
the general list.

Q: What is the anticipated release schedule?

Mid-May

Q: What components will be included in OFED and how is this decided?

- Components will include:

 - HCA driver - mthca
 - HCA driver - ipath
 - Core
 - IPoIB
 - SDP
 - RDS
 - SRP initiator
 - iSER initiator
 - OSU MPI
 - Open MPI
 - uDAPL
 - OpenSM
 - Diagnostic tools
 - Performance tests

The decision to include components is based on customer demand and level
of robustness and stability.  Components will be categorized in one of
three ways: 

  1) Basic: GA components that installed in a typical installation.

  2) Add-on: Components that can be installed optionally.

  3) Technology preview: Components where quality level is not GA, but
can be used by customers for technology development.

Q: When bugs are found, how will they be fixed?

- Fixes to release candidates are coordinated by the OFED release
coordinator and maintainers in a controlled fashion. Each bug found is
first fixed on the trunk, and then merged into the release branch.

- Patches will be made available as RC updates, and fed back to
OpenFabrics SVN continuously.

- Availability of patches will not be gated by acceptance of patches
into OpenFabrics SVN.

- Urgent bug fixes can be directly delivered to customers by the distros
or vendors, but are rolled into a standard release as quickly as
possible.  The goal is to ensure that fixes are standardized and make it
to the next general release.