Can we increase our MTU? This would decrease the CPU load.

What HoopyCat means, I think, is that you should back up your claim that increasing the MTU would decrease CPU load. The CPU savings from doing this are probably so small as to be unmeasurable.

Wikipedia says so in the second paragraph.

*login edit edit commit*
Now it says that increasing MTU will summon flying pink unicorns.
Okay, not really, but... don't trust Wikipedia.

Sure, if you're routing millions of packets per second, or are running on an embedded device with limited performance, then it might make a difference. On what is effectively a quad xeon server, any difference would be below the margin of error.

The only real reliable way is to use an increased MTU (without unicorns) is over the internal network. This I suspect would give benefit for for database connections, proxies, drbd, load balancers.

Nodebalancer to backend seems like a place where jumbo frames could be used if the Nodebalancer assembles TCP packets from/to web clients gives the backend less to assemble. Only way to see what impact is by testing.

Some small stats/recommendations I could find were;

* An Oracle test Seems to indicate a throughput increase of %4. I'd assume other DB based protocols would suffer the same bottlenecks.
* DRBD documentation suggests a performance increase under a number of likely deployment scenarios.
* MySQL cluster