Im comming right from the VM-Ware ESX3.5 - Fedora8 - Zimbra 5.0.14 hell.
This combination was running really bad for weeks! I did everything i could to try to get the performance up in zimbra following the guides for large deployments wich is probably not the best source because we are only 8 users on a machine with 4 cpu xenon cores each running at 2000Mhz and 4GB Ram.
Last night i did (just for fun) tweak zimbra in the other way so i set:
mailboxd_java_heap_memory_percent = 10
mysql_memory_percent = 10 (yes i did change it in my.cnf)
instead of the default 30 and 40 percent!
And badaboom now everything works great!

Zimbra has great performance and is stable?!? In my world lowering the memory should have slowed zimbra down not speed up performance?!? Whats the deal? Whats happening?
... and NO! - memory did not swap before nor does it swap now! There was always enough free memory available to zimbra!
Before the large java process took about 27% of memory now it consumes only about 8% and everything runs better now?!?
Could someone come up with some explanation for this phenomenon?!? I get a little nervous about this really untypical behaviour!
(By the way what is the "soap_max_in_memory_buffer_size" setting for and shoul i set it to something else than zero?)
--- Offtopic but it may help some people :
If someone is running ESX 3.5 on xeons together with fedora 7 or 8 you should not turn paravirtualization on but off and use follwoing kernel parameters in /etc/grub.conf:
kernel /vmlinuz-2.6.26.8-57.fc8 ro root=/dev/system/lvroot clocksource=acpi_pm nohz=off highres=off
And throw away those evil vm-ware tools if you dont want your kernel swapping all the time and probably set vm.swapiness to zero.