[ofw] OFA Windows Working Group 4/26 meeting minutes

[Smith, Stan]

Attendees:
---------
Microsoft
Mellanox
Intel

Agenda:
-------

==Reports from the OFA International Workshop in Monterey.

[Eric Lantz - microsoft.com]

Many good hallway conversations as in past years.
Seemingly reduced attendance.
Technical presentation content increased although not enough for all participants.
Ultra Fast sockets (beyond IPoIB and SDP), async operation highly desirable.
MPI wants to be notified by OpenSM of HCA/node failure; fabric status integration.
Support for multiple and 'larger' subnets.
OpenSM and diagnostics should be integrated with existing enterprise management frameworks.
Executive meeting discussions on RDMA training, Linux only to date; Door is open for Windows training - Who/how?
GPU support discussed, no movement for winOFED stack on GPU.

== winOFED 3.0 release discussion

target late Q2 functionality freeze.
target Q3 winOFED 3.0 GA release.

OpenSM 3.3.9 will be included in the 3.0 release.

IPoIB with Connected mode will be included.
  Today's UD mode is the IPoIB default mode, Connected Mode (CM) can be enabled from the IPoIB device advanced properties page.
  Stan will deliver a script suitable for clusrun execution which enables/disables IPoIB Connected Mode.

Network direct V2 provider support is questionable for 3.0 release - Eric will update ASAP.

== ROCE support

Mid-May to June timeframe for ROCE code injection into winOFED SVN source tree.
Mellanox noted that ROCE support is non-functional unless a proprietary Mellanox Ethernet driver is also loaded with winOFED stack.
Mellanox suggests the binary only, proprietary Ethernet driver be included in winOFED SVN tree and installed with winOFED.
Ensuing discussion followed - no resolution so far - still digesting information.

ROCE over Connect-X2 HCA is only supported on Server 2008 R2 and Windows 7.

Mellanox further clarified that no new development will occur for Server 2003 and XP InfiniHost HCA drivers.
Bug fixing will continue.

WWG is considering not supplying winOFED installers for Server 2003, XP and IA64 - any heartburn?

Uri noted Windows OpenSM lags behind OFED for Linux OpenSM w.r.t. bug fixes.
Stan suggested that Mellanox again become the OpenSM maintainer, as Stan became involved with OpenSM only because no one else was picking up the task.
If this is not possible, then we live with the cycles Stan can direct towards OpenSM maintenance.

Uri noted problems in supporting winmad/winverbs.
The WWG welcomed Uri to the Open Source development process; in that all code is open for review, as are all patches.
If you find a bug, then suggest a patch; all patches are subject to review and modifications by the maintainer.
Eric Lantz identified the 7 + Mellanox Windows team winOFED developers in response to the question as to why OFED for Linux had more maintainers.

We all agreed to work together towards identifying which modules need attention and to figure out how to apply engineering attention.

Eric inquired about the winOFED management interfaces we discussed at the last WWG meeting.
Sean suggested the management interfaces in libibnetdiscover (trunk\ulp\libibnetdisc\*) as candidates to which the MS HPC mgmt interface might hook into.
Sean noted that many OFED for Linux diagnostic tools are actually Perl scripts which process output from other OFED diag tools.
Best approach would be for all OFED diagnostic tools to be executables as including Perl in the winOFED distribution is not a good idea; perhaps PowerShell?