[LITMUS^RT] Question about SEND_RESCHED overheads

[Björn Brandenburg]

> On 02 Oct 2015, at 16:34, Namhoon Kim <namhoonk at cs.unc.edu> wrote:
> 
> According to this log, the maximum value (SEND_RESCHED_END - SEND_RESCHED_START) is 1151758398 (except the overflow values) cycles. It is 1.439 sec. This value doesn't match with the value shown in kernelshark. 

Sounds like you have non-synchornized, per-processor cycle counters that do not share a common time zero, or that are subject to substantial drift.

> Also, the timestamp of seq. 5 is earlier than the previous timestamp. This happens very often. 
> 
> Is this the right way to measure IPI overheads (SEND_RESCHED_END - SEND_RESCHED_START and ftcat commands)?

Yes.

> Does feather-trace use different timestamp for ft_cpu_traceX and ft_msg_traceX?

Have you checked the source code?

> Because the other overheads (e.g., SCHED, CXS, TICKS, RELEASE) are measured correctly.

The other overheads are measured with a single clock source.

> I also ran the same tasks on kvm machine, and it does not have any problems. This problem happens only on ARM machines. 

You probably have a non-synchronized, per-CPU clock source on your ARM board.

- Björn