On Tue, 21 Aug 2012, Lars Marowsky-Bree wrote: > On 2012-08-21T13:16:29, David Lang <[email protected]> wrote: > >> with ldirectord you have an extra network hop, and you have all your >> traffic going through one system. This is a scalability bottleneck as >> well as bing a separate system to configure. >> >> CLUSTERIP isn't the solution to every problem, but it works really >> well for many problems. > > To complete the list of issues though is that with clusterip, every > participating node sees all inbound traffic, and discards everything but > its own (CPU load). > > There's always a trade-off with everything.
agreed. CLUSTERIP also has an additional utility that if you use it with a one-way protocol (for example UDP syslog), you can have multiple sets of systems using the same IP address. This can make it so that the sender only needs to send one copy of the data instead of the sender needing to send one copy to each set of recievers. I've got a case where this actually avoids network bottlenecks on the sending side. >> It's especially good as a simple migration from active/passive. You >> just add a third box to the cluster and change the configuration so >> that instead of your VIP moving from one box to another, it's instead >> shared between the systems. This is a much smaller step than setting >> up an external load balancer system. > > I'm not quite sure what you need the 3rd box for? Of course, an odd > number of nodes is always a good idea, but not required by clusterip. if you need two boxes to handle the load, you should configure your cluster with at least three boxes so that if one box fails you can still handle the load with the remaining systems. David Lang _______________________________________________ Linux-HA mailing list [email protected] http://lists.linux-ha.org/mailman/listinfo/linux-ha See also: http://linux-ha.org/ReportingProblems
