On Wed, Jan 24, 2018 at 09:17:55AM +0100, Christoffer Dall wrote:
> Hi Shih-Wei,
>
> On Fri, Jan 19, 2018 at 04:57:55PM -0500, Shih-Wei Li wrote:
> > Thanks for the feedback about the mistakes in math and some issues in
> > naming, print msg, and coding style. I'll be careful and try to avoid
> > the same problems the next patch set. Sorry for all of the confusion.
> >
> > So we now skip the test when "sample == 0" happens over 1000 times.
> > This is only due to the case that "cost is < 1/cntfrq" since it's not
> > possible for the tick to overflow for that many times. Did I miss
> > something here? I do agree that we should output better msgs to tell
> > users that the cost of a certain test is constantly smaller than a
> > tick.
> >
>
> I think for things like vmexit counts, it's very likely that all the
> samples will result in 0 ticks on many systems (fast CPU and slow arch
> counter; the architecture doesn't give us any guarantees here). For
> example, a system with a 2 GHz CPU and a 1 MHz counter will give you a
> granularity of 2000 cycles for each counter tick, which is not that
> useful for low-level tuning of KVM.
>
> So what I thought we were going to do was:
>
> main_test_function()
> {
> long ntimes = NTIMES;
> long cost, total_cost;
>
> cnt1 = read_cnt();
> do {
> run_test();
> } while(ntimes--);
> cnt2 = read_cnt();
>
> if (verify_sane_counter(cnt1, cnt2))
> return;
>
> total_cost = to_nanoseconds(cnt2 - cnt1);
> cost = total_cost / NTIMES;
> printf("testname: %l (%l)\n", cost, total_cost);
> }
>
> And in that way amortize the potential lack of precision over all the
> iterations. Did I miss some prior discussion about why that was a bad
> idea?
Not that I know of, but I missed the above proposal, which I completely
agree with.
>
> It would also be possible to have two functions, one that does the above
> and one that does a per-run measurement, in case the user wants to know
> min/max/stddev and is running on a system with sufficient precision.
> The method could be chosen via an argument.
We should be able to just make NTIMES variable, setting it to one when a
per-run measurement is desired, right? Anyway, I'm fine with having set
to a reasonable value, not variable / no argument, for initial merge.
Thanks,
drew
