[ewg] Infiniband Interoperability

[Richard Croucher]

The InfiniBand fabric  knows very little about IpoIB, it is handled by the
host OS stack, however it does need capabilities such as multicast to work
properly for ARP name resolution.

 

The problem you describe sounds similar to a situation I encountered running
multiple, incompatible SM's.

Make sure you only have a single vendor SM.   Whilst the OFED SM build is
fine, I have found many vendors hack their distro's so they either ignore or
always win the SM election.   Explicitly disable all SM's on all
environments you don't want to be running. Don't rely on SM priority across
different implementations .  I'd recommend running openSM on Linux and
disabling all others.

 

Identifying that no SM running is easy, since the ports don't get LIDs,
however when multiple SM's are running it sort of works, since the different
SM's discover which LIDS have been allocated when they scan the fabric.  The
problem I saw was with multicast, each SM had its independent  and different
view of the MC nodes and paths.

 

Sun did their own InfiniBand stack implementation, including SM and is
completely independent of OFED.   I used it a few years ago and  IPoIB
interoperated fine with Linux.

 

You don't say which InfiniBand distro you are running on windows.

 

Linux IPoIB defaults to Connected mode.  This is an optional feature of
IpoIB.  You may want to try setting it to the mandatory datagram mode in
each of your environments. You can disable CM mode in openib.conf on the
linux nodes

 

From: ewg-bounces at lists.openfabrics.org
[mailto:ewg-bounces at lists.openfabrics.org] On Behalf Of Jeff Becker
Sent: 29 June 2010 19:26
To: Matt Breitbach; ewg at lists.openfabrics.org
Subject: Re: [ewg] Infiniband Interoperability

 

Hi Matt

On 06/29/10 10:15, Matt Breitbach wrote: 

So I know that this message isn't about starting a new group.  I've actually
tried to join one of the mailing lists but it failed to sign me up.
  

You probably tried to sign up for general at lists.openfabrics.org, which
doesn't really exist anymore (although the archives are still there for
searching). It sounds like you should post your question on
ewg at lists.openfabrics.org. I'll forward it for you.

Jeff Becker
OpenFabrics Server admin




 
I'm working on getting an InfiniBand setup working in a mixed environment of
Windows, Linux, and OpenSolaris.  I'm having huge difficulties getting the
IB network to be stable.  We're working with IPoIB mainly, and get drops
between the Windows and Linux/OpenSolaris systems.
 
We had a professional take a look at our configuration, and he thought that
OpenSM was configured properly, and actually had the network stable for
about 2 days, but it degraded severely after that.  We are now to the point
that most times the OpenSolaris box is unreachable from the Windows systems,
and sometimes from the Linux system (which is running OpenSM).
 
Is there any direction that you would be able to point me in to get some
advice on this or some high-end consulting?  We've invested nearly 3 months
into this project only to have a bladecenter with 4 IB Capable blades and a
SuperMicro 4U server that can't reliably communicate over the InfiniBand
network.
 
-Matt Breitbach
zfsbuild.com (our storage writeup about OpenSolaris ZFS over InfiniBand)
 
 
  

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ewg/attachments/20100630/ad332b4e/attachment.html>