refperf: Add warmup and cooldown processing phases
authorPaul E. McKenney <paulmck@kernel.org>
Tue, 26 May 2020 19:34:57 +0000 (12:34 -0700)
committerPaul E. McKenney <paulmck@kernel.org>
Mon, 29 Jun 2020 19:00:45 +0000 (12:00 -0700)
This commit causes all the readers to start running unmeasured load
until all readers have done at least one such run (thus having warmed
up), then run the measured load, and then run unmeasured load until all
readers have completed their measured load.  This approach avoids any
thread running measured load while other readers are idle.

Cc: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
kernel/rcu/refperf.c

index 234bb0e84a8b9d71418da197dd56221a005b0132..445190b97b052580907eef6635a2affd7b758add 100644 (file)
@@ -100,6 +100,8 @@ static atomic_t nreaders_exp;
 // Use to wait for all threads to start.
 static atomic_t n_init;
 static atomic_t n_started;
+static atomic_t n_warmedup;
+static atomic_t n_cooleddown;
 
 // Track which experiment is currently running.
 static int exp_idx;
@@ -260,8 +262,15 @@ repeat:
 
        VERBOSE_PERFOUT("ref_perf_reader %ld: experiment %d started", me, exp_idx);
 
-       // To prevent noise, keep interrupts disabled. This also has the
-       // effect of preventing entries into slow path for rcu_read_unlock().
+
+       // To reduce noise, do an initial cache-warming invocation, check
+       // in, and then keep warming until everyone has checked in.
+       cur_ops->readsection(loops);
+       if (!atomic_dec_return(&n_warmedup))
+               while (atomic_read_acquire(&n_warmedup))
+                       cur_ops->readsection(loops);
+       // Also keep interrupts disabled.  This also has the effect
+       // of preventing entries into slow path for rcu_read_unlock().
        local_irq_save(flags);
        start = ktime_get_mono_fast_ns();
 
@@ -271,6 +280,11 @@ repeat:
        local_irq_restore(flags);
 
        rt->last_duration_ns = WARN_ON_ONCE(duration < 0) ? 0 : duration;
+       // To reduce runtime-skew noise, do maintain-load invocations until
+       // everyone is done.
+       if (!atomic_dec_return(&n_cooleddown))
+               while (atomic_read_acquire(&n_cooleddown))
+                       cur_ops->readsection(loops);
 
        if (atomic_dec_and_test(&nreaders_exp))
                wake_up(&main_wq);
@@ -372,6 +386,8 @@ static int main_func(void *arg)
                reset_readers();
                atomic_set(&nreaders_exp, nreaders);
                atomic_set(&n_started, nreaders);
+               atomic_set(&n_warmedup, nreaders);
+               atomic_set(&n_cooleddown, nreaders);
 
                exp_idx = exp;