How does microbilling work?
Microbilling refers to the accounting for the CPU time that is used by a thread to a much finer resolution than the clock period between tick interrupts.
The thread scheduler has been implemented where threads send or receive messages many times (as opposed to a single time) per clock period. Adaptive partitioning scheduling would not be possible if we were limited to counting integer ticks of CPU time. That's because most threads send or receive messages, or otherwise block, many times per clock period.
Microbilling works by taking a fine-resolution timestamp every time a thread changes state from ready to not-ready, and charging differences between sequential timestamps against that partition's used CPU cycles count.
Microbilling uses the system call ClockCycles() to get that fine-resolution timestamp.
How often does thread scheduler microbill?
The thread scheduler microbills each time that:
How does ClockCycles() work?
The thread scheduler always depends on the processor being used. On x86 processors, Neutrino uses a free-running counter that is implemented on the CPU chip itself. This counter is read with a single instruction.
On PowerPC targets, Neutrino reads a similar free-running counter with just a few instructions. In these situations, ClockCycles() increments typically at about the processor's clock rates (i.e. ClockCycles() increases by 3 billion counts every second on a 3Ghz machine).
On both x86 and PowerPC processors, ClockCyles() increase by about 1 billion counts every second on a 1 GHz processor.
On processors that don't have a free-running counter for the purpose of being a fine-grained clock, Neutrino emulates ClockCyles(). For example, on ARM processors, Neutrino reads the intermediate value of the countdown timer that's used to trigger the clock interrupts. This value tells how far you're into the current clock tick. Neutrino further adds a scaled version of how far you're into the current clock tick to a constant determined at the last clock tick to get an emulated ClockCycles() value.
On some processors, such as ARM, the countdown timer used for emulating ClockCycles() is located off-chip and requires slow I/O operations to read it. On other processors, such as MIPS, the countdown timer is located on-chip, and can be quickly read.
How accurate is microbilling?
See the next answer.
How accurate is ClockCycles()?
The accuracy of microbilling or ClockCycles() is determined by the accuracy of the clock oscillator source used in the CPU. However, since the scheduling is relative between partitions, it doesn't require ClockCycles() be equal to the absolute time; it requires only that ClockCycles() be proportional to the work done by CPU. In fact, a wrongly calibrated ClockCycles() has no effect on the accuracy of the thread scheduler.
What is the resolution of thread timing?
It's the resolution of the ClockCycles() function. The resolution of clock cycles varies from platform to platform. In most cases, the resolution is much finer.