Re: [Condor-users] Default setting of NETWORK_INTERFACE

[Dan Bradley <dan@xxxxxxxxxxxx>]

Stuart,

The algorithm for picking an interface turns out to be a little more 
complicated than just picking the "first one found".  I will correct 
that in the manual.

The algorithm (as of my reading in 6.7.19 source code) is this:

Loop through DNS host records returned by gethostbyname() for the 
current host.  Pick one of the IP addresses from this list.  The first 
preference is an IP address other than anything matching the following 
addresses 10.*, 172.16.*, 192.168.*, 127.0.0.1.  The second choice is 
anything not matching 127.0.0.1.  The third choice is 127.0.0.1.  If 
there are multiple options of equal preference, the first one returned 
by gethostbyname() is used.  (However, I just discovered that the 
reverse is true for the private addresses 10.*, 172.16.*, etc.  In that 
case the _last_ one in the list is returned.  This is probably not 
intentional, so we'll review that.)

The new behavior added in 6.7.18 is that if you specify 
BIND_ALL_INTERFACES=True, then the choice of IP address to advertise is 
dynamic, depending on what network interface is actually used when 
communicating with the collector.  This allows, for example, a 
multi-homed schedd to flock jobs to two different sites in separate 
networks.  It also allows glidein startds that land on a multi-homed 
machine to more dependably choose the correct network interface to 
advertise so that they are reachable from the outside.

--Dan


On May 2, 2006, at 5:27 PM, Stuart Anderson wrote:

> 	On a multi-homed Linux machine what is the definition of
> "first interface found" for the default setting of NETWORK_INTERFACE
> in version 6.7.18? For example, is it the first one returned by
> /sbin/ifconfig, possibly excluding "lo"? Is there any ambiguity when
> virtual interfaces are configured?
>
> 	The default is currently working for me, but I am curious whether
> I am just lucky and should be explicitly setting NETWORK_INTERFACE or
> if the default algorithm is deterministic and appropriate to my setup.
>
> Thanks.
>
> --
> Stuart Anderson  anderson@xxxxxxxxxxxxxxxx
> http://www.ligo.caltech.edu/~anderson
> _______________________________________________
> Condor-users mailing list
> Condor-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/condor-users