On the time slice induced accuracy

[KarstenStahl]

Hello,

In [1] it is described that “the timeslice must be set significantly shorter than the shortest delay that is requested by any model.” I know that scheduled events are always executed on timeslice boundaries and the duration for the next timeslice is always adapted to the shortest planned events [1]. Therefore, I am confused about the part “significantly shorter”. In my understanding no timing inaccuracies should occur as long the time slice is <= “the shortest delay that is requested by any model”. At any time if an event is scheduled in a time slice it is then ensured that the event gets scheduled in the next time slice which can be adapted.

I created a platform where the CPU polls results from a custom peripheral. The peripheral sets the internal result flag to ready after a scheduled event (820ns) after a memory access which starts a calculation. Noticeable accuracy errors were only seen after a time slice bigger than 820ns was selected. However, even before that, small inaccuracies occurred. For instance, I configured a MIPS value of 8.383 which corresponds to an instruction length of 119.28ns per instruction. The quantum size was configured to be 0.00000001s (command line) and no further delays were used in the simulation. However, around 1% of instruction required 238ns instead of the configured 119ns. The timings per instruction were measured with the

[DuncGrah]

for (1) I think you are correct. The time slice must be smaller than the minimum delay required. I think this requires a minor documentation change.

for (2) I cannot think of an obvious reason that this may happen. Is it consistent i.e. always the same instructions that take longer when doing multiple runs?
I wonder if you can provide your test csase to us to run here and look at. If you can please PM me to let me know how to obtain it.

Thanks