[ofw] WWG 11-29-10 meeting minutes

[Smith, Stan]

OpenFabrics Alliance / Windows Working Group

Attendees
---------
Rupert Dance (lamprey networks)
Sean Hefty (Intel)
Eric Lantz (Microsoft)
Tzachi Dar (Mellanox)
Leonid Keller (Mellanox)
Stan Smith (Intel)

winOFED 2.3 Release Status
--------------------------

RC7 target release is Friday 12-3-10; possibly before.
winOFED 2.3 GA target is Friday 12-10-10.
Numerous IPoIB fixes, focus validation on IPoIB.
Microsoft will start a validation cycle as soon as RC7 is released.
 MS validation includes MSMPI + commercial applications.

Mellanox status
---------------
ROCE patches coming to a theater near you... HCA, ND providers: ibal, winverbs + V2.
HP platform requires ConnectX (mlx4) patches.
GRH (Global Route Header) support impacts ipoib, rdma_cm.

NetworkDirect Status
--------------------
The V2 provider is coming along now. Fab has prototypes of some of the ND logo tests, as well as MSMPI.  Working out issues in MSMPI currently when running over V1 providers, will shift to fixing up the V2 provider when he has things working.

ND tests need to be included in SVN src tree - Eric agrees.
V2 provider will be a shim over V1 interfaces initially.

Bing Maps housed in MS 'shipping container' data centers have an IB fabric within each container and between containers.

IB Diagnostics
--------------
MS has expressed a strong desire for IB diags which give an HPC cluster-wide view. Full spectrum from HCA firmware consistency to real-time data rate analysis.
MS HPC will update HPC diags in Q1'11 after Server 2008 R2 (SP1) is released; R2 sp1 target is very soon.
MS HPC diags currently parse vstat output in addition to WMI database. Question asked if raw text parsing is good enough or is an API needed.
Tzachi noted an API is safer for the long run. Sean noted raw text parsing can grow to be very slow as text volume increases.
MS laments current IB diags are single-system centric, lacking a cluster wide perspective.

Opens
-----
Tzachi requested a timely review of Irena's reference counting debug plan recently submitted to the list.

Tzachi also requested more due diligence on patch testing before SVN commits.

IPoIB Performance is now a Mellanox priority:
  Rx flow is ripe for performance enhancements.
  NDIS RSS (Receive Side Scaling) support under consideration, MS can be available for consultation.
  Core affinity
  Interrupt redirection.

Rupert noted some nomenclature confusion between ROCE (RDMA over Converged Ethernet) and IBoE (Infiniband over Ethernet).
Pick one or the other, do not use both.
Sean noted ROCE is an IBTA defined name.

stan.