Delayed ACK alongside DHCP failover

[John Wobus]

On Nov 21, 2014, at 10:51 AM, Terry Burton wrote:
> An ISC DHCP IO performance article [1] recommends considering the
> delayed ACK feature prior to the sledgehammer approach of
> dont-use-fsync.

We never used delayed ACK or "skip the fsyncs"
so I can't report any production experience.

But we've had to deal with performance.
We chose not to go with delayed ACK for the same
reasons you are worrying with: FUD from very old
reports and we weren't able to find reports of
much recent experience.  In trial, we experimented
with ramdisk as well as using a controller with writeback
cache and the latter worked well for us
performance-wise and we went with that.  Apparently
the current common writeback controller technology is a
volatile ram, a capacitor providing power to "finish up"
when power fails, and some flash to use only during
power failures.

I would have no trouble using ramdisk along with our own
periodic lease-file copying (e.g. minutely), but then we
already run our system around hour-long leases, so in the worst
case (which we've never seen), our troubles would be over
within an hour.  However, writeback cache has worked so well
that we've gone that way: better numbers in test than
ramdisk.  What we really need now is "instant daemon startup".

John Wobus
Cornell U IT