Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Sebastien Roy <sebastien.roy@delphix.com>
Reviewed by: Igor Kozhukhov <igor@dilos.org>
The APIC is used as a timer in Illumos. Specifically, it is used by the
callout and cyclic frameworks to generate an interrupt around the time that
the closest timer would expire. Once in the interrupt context those
frameworks call `gethrtime()` to determine which timers have expired, thus
the system doesn't solely rely on the accuracy of the APIC.
If the APIC is lagging behind the real time then we will have more jitter
and shorter timeouts will tend to be late. If the APIC is quicker than it
should then we will generate an excessive amount of interrupts as the APIC
would fire an interrupt before any timers expire. In any case, I've tested
what happens if the APIC is severely miscalibrated (10% or 1000% of target
speed) and it doesn't seem to create any unstability on the system.
With 1000% of the speed: we'd see a significant increase of the number of
interrupts fired, especially when system is idle:
CPU minf mjf xcal intr ithr csw icsw migr smtx srw syscl usr sys dt idl
0 41 0 5 9711 247 343 6 20 3 0 527 1 3 0 96
1 79 0 14 9366 409 1046 8 20 4 0 2894 1 3 0 96
vs, normally:
CPU minf mjf xcal intr ithr csw icsw migr smtx srw syscl usr sys dt idl
0 120 0 10 797 254 1082 9 20 3 0 2564 1 2 0 97
1 80 0 11 830 387 385 7 19 4 0 1175 1 1 0 98
The way that the APIC is calibrated is by using the 8254 fixed frequency
timer (PIT). We wait for it to count a certain amount of ticks and then we
check how many ticks does the APIC count in the same time interval. The main
issue is that on some hypervisors, notably hyperv, both the 8254 and the
APIC are emulated and thus can sometimes be inconsistent.
I've done an experiment to measure how much of an effect do those
inconsistencies have on the apic calibration factor (which determines how
many apic ticks pass in a certain amount of nanoseconds), and here are the
results for about 15000 measurements (done by performing 1000 measurements
at a time on each boot).
The main observation is that calibration doesn't seem to change from boot to
boot and that the accuracy of measurements doesn't seem to have any
correlation to the given time of measurement, which means that very
inaccurate measurements happen randomly. Most measurements are quite
accurate, except for some rare outliers (as can be seen in the graph). It
was determined that a 5-value median filter would significantly reduce the
worst case calibrations.
In the results below, `stdev %` is the standard deviation divided by the
average; `min %` is how far is the lowest calibration value measured
compared to the average and `max %` is how far is the highest calibration
value measured to the average.
Base Results:
stdev % min % max %
dcenter 0.02 0.2 0.2
AWS 0.02 1.4 0.1
hyperv 0.79 6.4 5.5
Azure 2.87 35.1 331.1
Using 5-value Median Filter:
stdev % min % max %
dcenter 0.01 0.02 0.04
AWS 0.01 0.01 0.03
hyperv 0.47 1.47 1.76
Azure 0.50 2.67 1.39
As we can see, using the median filter significantly reduces the worst-case
(min/max) miscalibrations on all platforms, and seems to be a necessity on
Azure to insure a proper worst-case calibration.
Closes openzfs#578