Issues with CPU Load Measurement

There is been support for CPU load measurement for a long time. Enabled with:

  CONFIG_SCHED_CPULOAD=y
  CONFIG_SCHED_CPULOAD_TIMECONSTANT=2

The main problem with the existing logic is that it ran synchronously with the system timer. That turns out to be useless for measurement of the performance of some tasks: Many tasks operate synchronously with the timer interrupt, i.e., timed event is the stimulus for task execution. Those tasks are always sampled at essentially the same point in the execution profile leading to nonsense results. In order to get believable results in all cases you would need to:

  • Sample at times are completely random with respect to program behavior, and
  • Sample at a high rate or for a very long time.

External Clock

In order to work around these things, an option to use an external clock was added:

  CONFIG_SCHED_CPULOAD_EXTCLK=y
  CONFIG_SCHED_CPULOAD_TICKSPERSEC=77

This should work well if CONFIG_SCHED_CPULOAD_TICKSPERSEC is prime and less than CONFIG_USEC_PER_TICK.
However, you were completely on your own one how to implement this external clock. But no longer.

Using a Oneshot Timer to Drive CPU Load Measurement

Recently, a generic, platform-independent one-shot lower half interface was developed. That interface is described in include/nuttx/timers/oneshot.h. There are implementations of the one shot timer lower half available for STM32, STM32L4, SAM4CM, SAMA5D3/4, and SAMV71/SAME70.

And now there is also an OS component that can be initialized and used to to drive the CPU load sampling from a precision, high rate oneshot timer.

The implementation of that feature resides at sched/sched/sched_cpuload_oneshot.c and is enabled with:

  CONFIG_ONESHOT=y
  CONFIG_CPULOAD_ONESHOT=y

Perhaps with additional configuration options that are likely required to enable board-specific oneshot timer lower-half support. The oneshot timer must still be configured by board specific logic which must call:

  void sched_oneshot_extclk(FAR struct oneshot_lowerhalf_s *lower);

To start the CPU load measurement.

sched_oneshot_extclk() is prototyped in include/nuttx/clock.h. There is some example setup code in the NuttX simulation code at boards/sim/sim/sim/src/sim_bringup.c.

Entropy

Another recent addition to NuttX came from David Alessio. David contributed support for /dev/urandom with a built-in XorShift128 pseduo-random number generator (PRNG). I have detached the XorShift128 implementation from the /dev/urandom implementation and moved it to /libc/misc so that is it available for other purposes.

In particular, there is an option to use the XorShift128 PRNG to add entropy to the CPU load measurement of the oneshot timer. That feature is enabled with:

  CONFIG_CPULOAD_ONESHOT_ENTROPY=n

for n=1-30 and disabled with CONFIG_CPULOAD_ONESHOT_ENTROPY=0. This value represents the number of bits of entropy that will be added to the oneshot interval delays. The oneshot timer will be set to the following interval each time the oneshot timer is restarted is CONFIG_CPULOAD_ONESHOT_ENTROPY:

  CPULOAD_ONESHOT_NOMINAL - (CPULOAD_ONESHOT_ENTROPY / 2) + error + nrand(CPULOAD_ONESHOT_ENTROPY)

Where

  • CPULOAD_ONESHOT_NOMINAL is CONFIG_SCHED_CPULOAD_TICKSPERSEC in units of microseconds.
  • CPULOAD_ONESHOT_ENTROPY is (1 << CONFIG_CPULOAD_ONESHOT_ENTROPY), and
  • error is an error value that is retained from interval to interval so that although individual intervals are randomized, the average will still be equal to CONFIG_SCHED_CPULOAD_TICKSPERSEC.

If CONFIG_CPULOAD_ONESHOT_ENTROPY=0, then the interval delay will always be equal to CPULOAD_ONESHOT_NOMINAL.

  • No labels