David Matlack <dmatlack@xxxxxxxxxx> writes:
On Fri, Sep 9, 2022 at 10:31 AM Ricardo Koller <ricarkol@xxxxxxxxxx>
wrote:
On Fri, Sep 09, 2022 at 10:26:10AM -0700, David Matlack wrote:
> On Fri, Sep 09, 2022 at 12:43:00PM +0000, Colton Lewis wrote:
> > Create the ability to randomize page access order with the -a
> > argument, including the possibility that the same pages may be hit
> > multiple times during an iteration or not at all.
> >
> > Population sets random access to false.
>
> Please make sure to also explain the why in addition to the what.
>
Will do.
> > @@ -57,7 +58,13 @@ void perf_test_guest_code(uint32_t vcpu_id)
> >
> > while (true) {
> > for (i = 0; i < pages; i++) {
> > - uint64_t addr = gva + (i * pta->guest_page_size);
> > + guest_random(&rand);
> > +
> > + if (pta->random_access)
> > + addr = gva + ((rand % pages) *
pta->guest_page_size);
> > + else
> > + addr = gva + (i * pta->guest_page_size);
> > +
> > guest_random(&rand);
>
> Is it on purpose use a separate random number for access offset and
> read/write?
>
It's because of the following, from
https://lore.kernel.org/kvm/YxDvVyFpMC9U3O25@xxxxxxxxxx/
I think addr and write_percent need two different random numbers.
Otherwise, you will end up with a situation where all addresses
where
(rnd_arr[i] % 100 < pta->write_percent) will get a write
(always).
Something like this:
012345678 <= address
wwwrrrwww
837561249 <= access order
I think the best way to fix this is to abstract the random number
reading into something like get_next_rand(), and use it twice per
iteration.
Makes sense. Depending on how many bits of randomness we need (e.g.
read/write only needs 7) we could still use one random number. But the
bit manipulation would probably more complex than just generating
another random number (which looks like a fairly cheap calculation).
Colton can you add a comment here to explain the subtlety?
Will do.
