Reducing IO load for system mirroring

[Matt Hersant]

I work in an environment that used rsync heavily for system mirroring.  
We've noticed that some of our machines are reaching their IO limit during 
the bi-hourly rsync run.  To my understanding, rsync transfers only the 
files that have changed, but must read every file from the disk to make that 
determination.  Is there a way to improve this performance?  Perhaps a way 
to cache a list of changed files to give to rsync on the bi-hourly run?  It 
believe I heard about a kernel option for this which is used in HA 
clustering.  Can someone offer any guidance?