On Wednesday 13 May 2009 06:04:38 John Cronin wrote: > Getting slightly off topic, but still somewhat relevant. > > Linux has many flavors of Ethernet bonding. To be sure, link aggregation > resulting in increased bandwidth is generally supported on a single switch. > However, Linux does have an active-passive bonding that is specifically > intended for HA solutions. AIX has a similar configuration with the > unfortunate name of EtherChannel Network Backup Interface - it does NOT > rely on Cisco EtherChannel to work. Both of these create a "virtual NIC" > that hides the complexity, making the interface group appear to be a single > NIC. You don't need a bunch of switch link aggregation magic (802.11ad or > EtherChannel) to implement active-passive NIC failover in this manner. > > In my experience, both Linux and AIX Ethernet bonding are easier to use > than Sun IPMP, and they also are far more reliable. I have a lot of > experience with all three of these, and in my opinion IPMP is the worst - I > have experienced many "false failures" with IPMP, and I have had to do a > bunch of silliness with static routes to make it work in certain > environments (prior to the new link based IPMP - but it has issues of its > own too). I wish Sun would add an active-passive capability to their new > link aggregation capability (dladm) that works across switches. If they > did that, they would have the same capabilities as Linux and AIX network > bonding, with similar ease of use. It should be fairly trivial to > implement. > > The one advantage that IPMP has in active-active mode (e.g. NOT link based) > is that it can detect IP connectivity issues (via ping - not just Ethernet > link detection) on all NICs in an interface group. However, it is usually > issues with the IP connectivity checking that cause all my problems with > IPMP, and I would gladly trade it for a simple link based virtual solution > that looks like a single link to me.
The linux bonding driver can do ARP test, in order to detect uplink failures. This is not a layer 3 check, but in most solutions and configurations, it is a suitbale replacment. > > I have never used Linux Ethernet bonding or AIX Etherchannel Network Backup > Interface for VCS heartbeats, but I am pretty certain they would both work > fine. That said, I am not sure they would provide any significant benefit > over a "traditional" VCS heartbeat network configuration, using the same > number of "real" NICs. The benefit is that with such bonding method, you can survive the failure scenario I've described in my first email :) It is a fact Symantec understands that, as they are trying to solve it internally in LLT :) > > -- > John Cronin _______________________________________________ Veritas-ha maillist - [email protected] http://mailman.eng.auburn.edu/mailman/listinfo/veritas-ha
