The priority of each thread is the sum of three priority values,
Each thread's base priority level is initialized to 0 and is intended
to be adjusted externally (such as via a setpriority
system call or “renice” operation.) The values for the
base priority must be constrained to lie in the range -256..256.
Each thread's dynamic priority level is adjusted to prioritize
I/O-bound threads and penalize CPU-bound ones. A thread that blocks
or yields voluntarily has its dynamic priority raised. A thread that
is stopped by a clock interrupt has its dynamic priority lowered.
The change made to dynamic priority is
SCHED_DYNPRI_DELTA, which is fixed at 16.
The dynamic priority is clamped to the range
SCHED_DYNPRI_MIN to SCHED_DYNPRI_MAX
(-256..256). The lower bound prevents totally CPU-bound threads
from being abused too badly. The upper bound keeps threads from
getting too much credit for doing tons of little I/Os.
Finally, each thread has a compensation priority level. This is cleared to zero when a thread is chosen to run, and incremented each time a thread is not chosen to run. This means that even extremely low-priority threads will eventually be chosen to run.
The change made to the compensation priority level (the compensation priority delta) is a variable. The normal value of this variable is 1. However, it could potentially be set to 0, which in combination with a low base priority would cause a thread to be run more or less only when nothing else is runnable. It could also be potentially set higher, which could help to ensure a fast response time on a busy system. One could conceivably adjust this value dynamically as well.
The compensation priority level is not clamped, but is limited in practice by the fact that a thread with a high compensation priority will eventually be run and lose the compensation priority.
The values picked for all these constants are somewhat arbitrary and could be tuned.