thanks for answering. I'd like to point out that for the VLAN scenario we use ARP monitoring, so that a dead VLAN wil be detected hopefully (as opposed to a dead link between NIC and switch).
Our load on the bond will be high-volume TCP data (backups), so maybe even some reordering is acceptable (we have 128GB RAM in the machines).
Is there any document on the rx timeout or reordering settings?
Finally: "Support" is Novell Support for SLES11 SP1.
>>> Jay Vosburgh <fubar@...> schrieb am 26.06.2012 um 18:55 in Nachricht
> Ulrich Windl <Ulrich.Windl@...> wrote:
> >In Documentation/networking/bonding.txt I can read:
> > The balance-rr, balance-xor and broadcast modes generally
> >require that the switch have the appropriate ports grouped together.
> >We use balance-rr, and I wonder why
> >1) if two hosts have the same number of interfaces in the bond, why
> >can't you use straight-throuch cables? We did, and it actually worked,
> >but support said, it's not supported without a switch in between
> I don't know who "support" is, but, yes, you can do this. The
> limitation is that the hosts cannot communicate with anyone other than
> each other unless they have additional network interfaces.
> >2) If multiple hosts are involved, and all hosts have the same number
> >of interfaces in their bond, why can't the corresponding interfaces be
> >put into VLANs? We use that configuration, too, but support said the
> >switch must be set for "etherchannel". As we use different switches for
> >the VLANs to avoid SPoFs, ethercannel across switches is not possible
> >with our firmware (probably a Cisco patent issue).
> This also works (and doesn't need the switch ports in
> etherchannel mode), but with the caveat that if any slave fails, lots of
> packets sent to the host with the failed slave will be dropped. This
> can be set up with VLANs or actual switches (as long as those switches
> are not connected).
> >So can the quote above be interpreted for the generic case, not meaning
> >that in specific other cases the configurations cited are still expected
> >to work?
> Either of the above may require adjustment to the slave devices'
> rx-usecs ethtool setting (ethtool -c / -C), and/or adjustment to the
> tcp_reordering sysctl. The balance-rr mode will pretty much always
> reorder packets, and adjustment of the parameters is needed to mitigate
> the impact on TCP. If you use an application protocol on UDP that
> cannot tolerate packet reordering, then balance-rr is not an appropriate
> -Jay Vosburgh, IBM Linux Technology Center, fubar@...