Unexpected Realtime Delay, but Good Latency Values

3 axis mill with a 5i25 and 7i77 on a ASRock Q1900B. 

This machine ran for years on Wheezy and 2.7 RTAI, but I wanted to modernize it so I installed a 23" touchscreen and updated to 2.8.2, Debian 10 PREEMPT-RT and LinuxCNC. 

After the upgrade I get the error Unexpected Realtime Delay on task 0 with period 200000 and DoLt not cleared from previous servo thread. Servo thread rate probably too fast.

If I adjust the servo rate to 250000 I do not get this issue. I get a following error when trying to home the machine if I keep the 200000 rate and ignore the errors or if I use the 250000 servo period. 

Here is a picture of the latency histogram after running for ~30 minutes with youtube and 3 x GLXgears. Latency values were higher without isolcpus (which I never needed to use on the previous setup).
 

Shouldn't these latency values allow me to run an even faster servo period than 200000? 

Attached is my .ini and .hal. 

I have also tried 2.8.2 on RTAI and this resulted in slightly lower latency values, but still the same errors.

Do you need to run a faster servo thread period than the normal 1 ms?

To be honest, I'm not sure. My drives are configured for torque control and on this page, linuxcnc.org/docs/html/config/pncconf.htm

it says:

if you are building a closed loop servo system that controls torque (current) rather then velocity (voltage) a faster rate would be better - something like 200000.

If slowing down the rate to 250000 is acceptable with torque control than I can just do that, but I wanted to avoid changing the servo period as it would require retuning the drives, no? Also, I would think that my latency numbers are good enough that I would be able to keep the 200000 period.

250000 (250 usec = 4 KHz) should be OK for torque mode drives
and should not require a major change in tuning

I appreciate the suggestion, but at the end of my first post I wrote: I have also tried 2.8.2 on RTAI and this resulted in slightly lower latency values, but still the same errors.

Don't my latency values imply that I could easily run the 200000 servo period without errors?

The latency test is pretty optimistic, as it does no actual I/O
If you want to get a realistic idea of where the time goes,
take a look at all the tmax values after running LinuxCNC
for a long time:

halcmd show param *.tmax

(these are in units of CPU clocks on X86)