Ifgrps adds some intelligence to IP's source address selection and outbound interface selection to allow IP traffic to flow on any of the interfaces. The Solaris IP implementation adds a cache entry per destination IP address, which can be viewed with the netstat command: % netstat -ra | grep UHA Ifgrps come into play when creating these cache entries. Use the ndd command to enable/disable ifgrps, as follows: ndd -set /dev/ip ip_enable_group_ifs 0 When a cache entry is created with ip_enable_group_ifs set to 1, a source address for this entry is selected and the outgoing interface is assigned based on this source address. This behavior is new for 2.6, and is part of the ifgrps functionality. The algorithm is: 1. If there's a cache entry, use it, and send traffic out that interface regardless of source address. 2. If there's no cache entry, check if there's a source address already in the datagram. If so, create a cache entry with the interface corresponding to the source address. (NEW IN 2.6) 3. With no entry and no source address, round-robin to one of the interfaces in the interface group that routing says to use. (NEW IN 2.6) If ip_enable_group_ifs is set to 0, the machine reverts to pre-2.6 behavior: 1. If there's a cache entry, use it, and send traffic out that interface regardless of source address. 2. If there's no cache entry, use the interface routing indicates (it'll always be just one) and set the cache entry to that. Note: There are two important limitations to what is described above when ifgrps is enabled: 1.For logical interfaces (le0:1, le0:2 etc) the load balancing does not take place and only the physical interface associated with the logical interfaces le0 is listed in the interface group list. 2.For systems with a number of physical interfaces and some logical interfaces, the load balancing takes place only on the physical interfaces and not on the logical interfaces. From patch 105786: ------------------ Synopsis: SunOS 5.6: /kernel/drv/ip patch BugID 4077132: ip_enable_group_ifs=0 doesn't pick default hostname The ip patch fix restores the correct behaviour of interface groups. The new code has the ip_enable_group_ifs disabled by default, and so it restores the old behaviour and the first physical interface acts as the *primary interface*.