Re: [dpdk-stable] [dpdk-dev] [PATCH v1 1/2] test/rwlock: add perf test case

[Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>]

> >
> > Hi,
> >
> > >
> > > Add performance test on all available cores to benchmark the scaling
> > > up performance and fairness of rw_lock.
> > >
> > > Fixes: af75078faf ("first public release")
> > > Cc: stable@dpdk.org
> > >
> > > Suggested-by: Gavin Hu <gavin.hu@arm.com>
> > > Signed-off-by: Joyce Kong <joyce.kong@arm.com>
> > > Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> > > Reviewed-by: Ola Liljedahl <ola.liljedahl@arm.com>
> > > Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> > > Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
> > > ---
> > >  test/test/test_rwlock.c | 71
> > > +++++++++++++++++++++++++++++++++++++++++++++++++
> > >  1 file changed, 71 insertions(+)
> > >
> > > diff --git a/test/test/test_rwlock.c b/test/test/test_rwlock.c index
> > > 29171c4..4766c09 100644
> > > --- a/test/test/test_rwlock.c
> > > +++ b/test/test/test_rwlock.c
> > > @@ -4,6 +4,7 @@
> > >
> > >  #include <stdio.h>
> > >  #include <stdint.h>
> > > +#include <inttypes.h>
> > >  #include <unistd.h>
> > >  #include <sys/queue.h>
> > >
> > > @@ -44,6 +45,7 @@
> > >
> > >  static rte_rwlock_t sl;
> > >  static rte_rwlock_t sl_tab[RTE_MAX_LCORE];
> > > +static rte_atomic32_t synchro;
> > >
> > >  static int
> > >  test_rwlock_per_core(__attribute__((unused)) void *arg) @@ -65,6
> > > +67,72 @@ test_rwlock_per_core(__attribute__((unused)) void *arg)
> > >  	return 0;
> > >  }
> > >
> > > +static rte_rwlock_t lk = RTE_RWLOCK_INITIALIZER; static uint64_t
> > > +lock_count[RTE_MAX_LCORE] = {0};
> > > +
> > > +#define TIME_MS 100
> > > +
> > > +static int
> > > +load_loop_fn(__attribute__((unused)) void *arg) {
> > > +	uint64_t time_diff = 0, begin;
> > > +	uint64_t hz = rte_get_timer_hz();
> > > +	uint64_t lcount = 0;
> > > +	const unsigned int lcore = rte_lcore_id();
> > > +
> > > +	/* wait synchro for slaves */
> > > +	if (lcore != rte_get_master_lcore())
> > > +		while (rte_atomic32_read(&synchro) == 0)
> > > +			;
> > > +
> > > +	begin = rte_rdtsc_precise();
> > > +	while (time_diff < hz * TIME_MS / 1000) {
> > > +		rte_rwlock_write_lock(&lk);
> > > +		rte_pause();
> >
> > Wouldn't it be more realistic to write/read some shared data here?
> > Again extra checking could be done in that case that lock behaves as
> > expected.
> Will do it in v2, thanks!
> >
> > > +		rte_rwlock_write_unlock(&lk);
> > > +		rte_rwlock_read_lock(&lk);
> > > +		rte_rwlock_read_lock(&lk);
> >
> > Wonder what is the point of double rdlock here?
> > Konstantin
> Double rd lock is to check rd locks will not block each other.
> Anyway I will remove it in v2 if no concerns here.
> >
> > > +		rte_pause();
> > > +		rte_rwlock_read_unlock(&lk);
> > > +		rte_rwlock_read_unlock(&lk);
> > > +		lcount++;
> > > +		/* delay to make lock duty cycle slightly realistic */
> > > +		rte_pause();
> > > +		time_diff = rte_rdtsc_precise() - begin;
> > > +	}
Should we change the way the measurement is done? We are measuring 'how many locks/unlocks per <certain time>'. This introduces more over head due to rte_rdtsc_precise call for every iteration. If we do, 'how many cycles it takes to do <certain number of locks/unlocks>', the over head of rte_rdtsc_precise can be amortized and will be very little.

> > > +	lock_count[lcore] = lcount;
> > > +	return 0;
> > > +}
> > > +