[CC Pablo] Hi Alex, On Wed, Aug 26, 2015 at 10:47:52AM -0700, Alex Gartrell wrote: > The configuration of ipvs at Facebook is relatively straightforward. All > ipvs instances bgp advertise a set of VIPs and the network prefers the > nearest one or uses ECMP in the event of a tie. For the uninitiated, ECMP > deterministically and statelessly load balances by hashing the packet > (usually a 5-tuple of protocol, saddr, daddr, sport, and dport) and using > that number as an index (basic hash table type logic). > > The problem is that ICMP packets (which contain really important > information like whether or not an MTU has been exceeded) will get a > different hash value and may end up at a different ipvs instance. With no > information about where to route these packets, they are dropped, creating > ICMP black holes and breaking Path MTU discovery. Suddenly, my mom's > pictures can't load and I'm fielding midday calls that I want nothing to do > with. > > To address this, this patch set introduces the ability to schedule icmp > packets which is gated by a sysctl net.ipv4.vs.schedule_icmp. If set to 0, > the old behavior is maintained -- otherwise ICMP packets are scheduled. Nice work. I have queued these up with Julian's Ack in the ipvs-next tree. At this stage my plan is to send a pull request to Pablo targeted at v4.4 after v4.3-rc1 has been released. I will probably rebase ipvs-next at that time. -- To unsubscribe from this list: send the line "unsubscribe lvs-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
