Re: [PATCH v2 1/2] lib/find_bit_benchmark: Add benchmark test for fns()

[Kuan-Wei Chiu <visitorckw@gmail.com>]

On Tue, Apr 30, 2024 at 10:24:03AM -0700, Yury Norov wrote:
> On Tue, Apr 30, 2024 at 01:49:11PM +0800, Kuan-Wei Chiu wrote:
> > Introduce a benchmark test for the fns(). It measures the total time
> > taken by fns() to process 1,000,000 test data generated using
> > get_random_long() for each n in the range [0, BITS_PER_LONG].
> 
> Can you also print an example of test output?
>  
> > Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
> > ---
> >  lib/find_bit_benchmark.c | 25 +++++++++++++++++++++++++
> >  1 file changed, 25 insertions(+)
> > 
> > diff --git a/lib/find_bit_benchmark.c b/lib/find_bit_benchmark.c
> > index d3fb09e6eff1..8712eacf3bbd 100644
> > --- a/lib/find_bit_benchmark.c
> > +++ b/lib/find_bit_benchmark.c
> > @@ -146,6 +146,28 @@ static int __init test_find_next_and_bit(const void *bitmap,
> >  	return 0;
> >  }
> >  
> > +static int __init test_fns(void)
> > +{
> > +	const unsigned long round = 1000000;
> > +	s64 time[BITS_PER_LONG + 1];
> > +	unsigned int i, n;
> > +	volatile unsigned long x, y;
> > +
> > +	for (n = 0; n <= BITS_PER_LONG; n++) {
> 
> n == BITS_PER_LONG is an error. Testing error case together with
> normal cases is even worse error because it fools readers.
>
My initial intention was to add a test for fns() always returning
BITS_PER_LONG. However, I agree that this is not a good idea and may
confuse readers.

> > +		time[n] = ktime_get();
> > +		for (i = 0; i < round; i++) {
> > +			x = get_random_long();
> > +			y = fns(x, n);
> > +		}
> 
> Here you count fns() + get_random_long() time. For your microbench
> purposes it would be better exclude a random number generation
> overhead.
> 
> > +		time[n] = ktime_get() - time[n];
> > +	}
> > +
> > +	for (n = 0; n <= BITS_PER_LONG; n++)
> > +		pr_err("fns: n = %2u: %12lld ns\n", n, time[n]);
> 
> Nah, not like that. Each test in there prints one line in the
> report. Let's keep it that way for test_fns() too. Unless we have
> a strong evidence that fns() for a particular input is worth to be
> tracked separately, let's just print a total gross?
> 
> > +
> > +	return 0;
> > +}
> 
> I'd suggest to modify it like:
> 
>         static unsigned long buf[1000000];
> 
>         static int __init test_fns(void)
>         {
>                 get_random_bytes(buf, ARRAY_SIZE(buf));

Instead of ARRAY_SIZE(buf), it should be sizeof(buf).

>                 time = ktime_get();
> 
>                 for (n = 0; n < BITS_PER_LONG; n++)
>                         for (i = 0; i < 1000000; i++)
>                                 fns(buf[i], n);
> 
>                 time = ktime_get() - time;
>                 pr_err(...);
>         }
>

That does seem like a better approach. I'll move it to lib/test_bitops
and send a v3 patch series.

Regards,
Kuan-Wei

> >  static int __init find_bit_test(void)
> >  {
> >  	unsigned long nbits = BITMAP_LEN / SPARSE;
> > @@ -186,6 +208,9 @@ static int __init find_bit_test(void)
> >  	test_find_first_and_bit(bitmap, bitmap2, BITMAP_LEN);
> >  	test_find_next_and_bit(bitmap, bitmap2, BITMAP_LEN);
> >  
> > +	pr_err("\nStart testing for fns()\n");
> > +	test_fns();
> 
> There are 2 sections in the test - one for regular, and another for
> sparse data. Adding a new section for a just one function doesn't look
> like a good idea.
> 
> Even more, the fns() is already tested here. Maybe test_bitops is a
> better place for this test?
> 
> > +
> >  	/*
> >  	 * Everything is OK. Return error just to let user run benchmark
> >  	 * again without annoying rmmod.
> > -- 
> > 2.34.1