Did find a performance bug in the core I don't remember much about (unfortunate choice of operations that flushed caches or something and realllly slowed down... something). Submitted the half-a-dozen line patch upstream that Joel Sherrill pulled without drama, even though my test case was overly complicated and opaque.
What we needed was something like Linux's scheduling queue disciplines [0]; it would have allowed us to say "here's a packet, schedule it for 10 mSec from now", and we'd be good as long as the vast majority of packets took less than 10 msec to wend through the stack. But that didn't exist at that point.
Because RTEMS is open-source and also because its internal structure is many orders of magnitude easier to understand than that of something like Linux, or even of some *BSD, extreme customization can be done with relatively modest work.
When you do not need them, it is easy to remove completely the implementations of many parts of the API and of many classes of kernel objects, to obtain a much leaner RTOS kernel.
Any causes of jitter from the kernel or device drivers (if they occur in higher priority tasks, as the lower priority tasks should not be able to interfere) should be relatively easy to identify and patch.
I have used once RTEMS for an application with many channels of synchronous multiplexed telephony, with very tough jitter requirements, and there have been no problems, but it was not RTEMS out-of-the-box, it was customized for the application, i.e. with about half of the kernel removed, a new interrupt handler added, and with custom device drivers.