[ewg] NODE_DESC=$(hostname -s) of openibd give localhost at boot

[Hal Rosenstock]

On 10/16/2014 3:19 AM, giggzounet wrote:
> Hi,
> 
> We have a small cluster with centos 6.5 and infiniband. OFED is
> installed and infiniband is working. The nodes get their IP with DHCP.

Is this using DHCP over IPoIB or something else ? Is it PXE boot ?

> At boot openibd is started without error. But the description of each
> node is "localhost", which is quite annoying. If I do a restart of
> openibd at the end of the boot, the description is set to the "correct"
> hostname (n01, n02,...).
> 
> I think it is a "problem" with the DHCP. At the start of openibd, the
> node has no IP. But I'm not sure.
> 
> Is it a classical problem? Is there a way to solve it?

It sounds like openibd is invoked on node before DHCP sets the IP
address so hostname -s returns "localhost" which openibd then uses to
put into the HCA's NodeDescription.

Below are 2 basic ways to deal with this:

One way to deal with this is to set the node description appropriately
after DHCP completes.

Another approach would be to delay the start of openibd until DHCP has
assigned the IP address but that won't be possible if it's using DHCP
over IPoIB.

There are some others but they're more "labor" intensive.

-- Hal

> Regards,
> Guillaume
> 
> _______________________________________________
> ewg mailing list
> ewg at lists.openfabrics.org
> http://lists.openfabrics.org/mailman/listinfo/ewg
>