Switch to frequency/throughput-based testing for test-cdeque

nhatminhle · nhatminhle · commit de55466df7e2 · 2012-08-09T12:12:09.000+02:00
diff --git a/libworkstream_df/benchmark-cdeque.sh b/libworkstream_df/benchmark-cdeque.sh
@@ -1,14 +1,18 @@
-breadth=3
-depth=15
+b=3
+d=15
+mu=$(awk "BEGIN{print ($b ^ $d) / 10}")
 nthread=2
-niter=40
-while getopts b:d:n:ps:t: opt; do
+niter=200
+while getopts b:d:f:n:pt: opt; do
 	case $opt in
 	b)
-		breadth=$OPTARG
+		b=$OPTARG
 		;;
 	d)
-		depth=$OPTARG
+		d=$OPTARG
+		;;
+	f)
+		mu=$OPTARG
 		;;
 	n)
 		nthread=$OPTARG
@@ -22,81 +26,25 @@ while getopts b:d:n:ps:t: opt; do
 done
 shift $((OPTIND - 1))
 
-tune()
-{
-	echo tuning $t $testargs -s $2 >&2
-	awk -f tune-test-cdeque.awk $1 "$t $testargs" $2 $nsteal $nthread
-}
-
-iter()
-{
-	for i in `seq 1 $1`; do
-		if [ $3 ]; then
-			farg="-f $3"
-		else
-			farg=
-		fi
-		$t $testargs -s $2 $farg | tail -n 1
-	done
-}
+tests='test-cdeque test-cdeque-nofences test-cdeque-c11 test-cdeque-dumbc11'
+testargs="-b $b -d $d -n $nthread"
 
-run()
-{
-	log=$t.$nthread.$b.$d.$1.log
-	f=$(tune 15 $1)
-	if [ $dry ]; then
-		echo $t $testargs -s $1 -f $f
-		$t $testargs -s $1 -f $f
-	else
-		echo $t $testargs -s $1 -f $f \> $log >&2
-		iter $niter $1 $f >$log
-	fi
+awk -v ni=$niter -v mu=$mu '
+BEGIN {
+	for (i = 1; i <= ni; ++i)
+		print -log(rand()) * mu
 }
+' >rand.$$
 
-do_tree()
-{
-	b=$1
-	d=$2
-	nsteal=$(awk -v b=$b -v d=$d '
-	BEGIN {
-		total = 0
-		if (b == 1)
-			total = d
-		else {
-			row = b
-			for (i = 0; i < d; ++i) {
-				row *= b
-				total += row
-			}
-		}
-		print total
-	}
-	')
-	echo jobs $nsteal >&2
-
-	tests='./test-cdeque ./test-cdeque-nofences ./test-cdeque-c11 ./test-cdeque-dumbc11'
-	testargs="-b $b -d $d -n $nthread"
-
-	for t in $tests; do
-		echo $t prerun >&2
-		[ $dry ] || iter 10 0 >/dev/null
-		f=0
-		run 0
-		s=1
-		n10=$((nsteal / 10))
-		while [ $s -lt $n10 ]; do
-			if [ $s -lt $((nthread * 10)) ]; then
-				echo skipping $t $testargs -s $s >&2
-			else
-				run $s
-			fi
-			s=$((s * 10))
-		done
-		for sratio10 in 1 2 4 6 8; do
-			s=$((sratio10 * nsteal / 10))
-			run $s
-		done
+for t in $tests; do
+	echo $t prerun >&2
+	for i in $(seq 1 10); do
+		[ $dry ] || ./$t $testargs -f 0 | tail -n 1 >/dev/null
 	done
-}
-
-do_tree $breadth $depth
+	log=freq.$t.$nthread.$b.$d.log
+	while read x; do
+		f=$(awk "BEGIN{print $x / ($nthread-1)}")
+		echo $t $testargs -f $f >&2 \>\> $log
+		[ $dry ] || ./$t $testargs -f $f | tail -n 1 >>$log
+	done <rand.$$
+done
diff --git a/libworkstream_df/plot-benchmark-cdeque.sh b/libworkstream_df/plot-benchmark-cdeque.sh
@@ -23,53 +23,54 @@ plot()
 {
 	gnuplot - <<EOF
 set terminal $type
+set xlabel '$4'
+set ylabel '$5'
 set output '$1.$n.$b.$d.$type'
-plot '$1.$n.$b.$d.data' title 'native' w lines, \\
-     '$1-nofences.$n.$b.$d.data' title 'nofences' w lines, \\
-     '$1-c11.$n.$b.$d.data' title 'c11' w lines, \\
-     '$1-dumbc11.$n.$b.$d.data' title 'seqcst' w lines
+set title '$3 (n=$n; b=$b; d=$d)'
+plot '$1.$n.$b.$d.data' title 'native' w $2, \\
+     '$1-nofences.$n.$b.$d.data' title 'nofences' w $2, \\
+     '$1-c11.$n.$b.$d.data' title 'c11' w $2, \\
+     '$1-dumbc11.$n.$b.$d.data' title 'seqcst' w $2
 set output '$1.$n.$b.$d.log.$type'
+set title '$3 (n=$n; b=$b; d=$d; log)'
 set log x
-plot '$1.$n.$b.$d.data' title 'native' w lines, \\
-     '$1-nofences.$n.$b.$d.data' title 'nofences' w lines, \\
-     '$1-c11.$n.$b.$d.data' title 'c11' w lines, \\
-     '$1-dumbc11.$n.$b.$d.data' title 'seqcst' w lines
+plot '$1.$n.$b.$d.data' title 'native' w $2, \\
+     '$1-nofences.$n.$b.$d.data' title 'nofences' w $2, \\
+     '$1-c11.$n.$b.$d.data' title 'c11' w $2, \\
+     '$1-dumbc11.$n.$b.$d.data' title 'seqcst' w $2
 EOF
 }
 
 tests='test-cdeque test-cdeque-nofences test-cdeque-c11 test-cdeque-dumbc11'
 
-for t in $tests; do
-	for log in $t.$n.$b.$d.*.log; do
-		s=${log%.log}
-		s=${s##*.}
-		awk "{x+=\$1;++n} END{print $s, x/n}" $log
-	done | sort -n >$t.$n.$b.$d.data
-done
-plot test-cdeque
-
-for t in $tests; do
-	for log in $t.$n.$b.$d.*.log; do
-		s=${log%.log}
-		s=${s##*.}
-		if [ $s -eq 0 ]; then
-			echo 0 1.0
-		else
-			awk "{x+=\$(NF-4)/\$NF;++n} END{print $s, x/n}" $log
-		fi
-	done | sort -n >steal.$t.$n.$b.$d.data
-done
-plot steal.test-cdeque
-
-for t in $tests; do
-	for log in $t.$n.$b.$d.*.log; do
-		s=${log%.log}
-		s=${s##*.}
-		if [ $s -eq 0 ]; then
-			echo 0 1.0
-		else
-			awk "{x+=\$(NF-4)/\$1;++n} END{print $s, x/n}" $log
-		fi
-	done | sort -n >stealtime.$t.$n.$b.$d.data
-done
-plot stealtime.test-cdeque
+if [ -n "$(find . -maxdepth 1 -name "test-*.$n.$b.$d.log")" ]; then
+	for t in $tests; do
+		awk '{print $NF, $1}' $t.$n.$b.$d.log |
+		sort -n >time.$t.$n.$b.$d.data
+	done
+	plot time.test-cdeque points		\
+	    'Execution time'				\
+	    'effective steals' 'execution time (s)'
+	for t in $tests; do
+		awk '{print $(NF-2), $1}' $t.$n.$b.$d.log |
+		sort -n >time_exp_tput.$t.$n.$b.$d.data
+	done
+	plot time_exp_tput.test-cdeque points			\
+	    'Execution time'						\
+	    'expected steal throughput (s⁻¹)' 'execution time (s)'
+	for t in $tests; do
+		awk '{print $NF/$1, $1}' $t.$n.$b.$d.log |
+		sort -n >time_eff_tput.$t.$n.$b.$d.data
+	done
+	plot time_eff_tput.test-cdeque points			\
+	    'Execution time'						\
+	    'effective steal throughput (s⁻¹)' 'execution time (s)'
+	for t in $tests; do
+		awk '{print $NF/$1, 2*$(NF-3)/$1}' $t.$n.$b.$d.log |
+		sort -n >tput_eff_tput.$t.$n.$b.$d.data
+	done
+	plot tput_eff_tput.test-cdeque points	\
+	    'Critical path throughput'			\
+	    'effective steal throughput (s⁻¹)'		\
+	    'push/take throughput (s⁻¹)'
+fi
diff --git a/libworkstream_df/stat-benchmark-cdeque.sh b/libworkstream_df/stat-benchmark-cdeque.sh
@@ -11,12 +11,12 @@ for log in test-cdeque*.log; do
 	mean=$(awk '{x+=$1;++n} END{print x/n}' $log)
 	stddev=$(awk -v m=$mean '{x+=($1-m)^2;++n} END{print sqrt(x/n)}' $log)
 	maxdev=$(awk -v m=$mean '{x=($1-m)^2} x>u{u=x} END{print sqrt(u)}' $log)
-	thief_avg=$(awk '{for(i=1;i<=NF-5;++i){x+=$i;++n}} END{print x/n}' $log)
+	thief_avg=$(awk '{for(i=1;i<=NF-6;++i){x+=$i;++n}} END{print x/n}' $log)
 	thief_avgerr90=$(awk						\
-	    '{for(i=1;i<=NF-5;++i){x+=$i/(0.9*$1);++n}} END{print x/n}'	\
+	    '{for(i=1;i<=NF-6;++i){x+=$i/(0.9*$1);++n}} END{print x/n}'	\
 	    $log)
 	steal_avg=$(awk '{x+=$(NF-1)==0?0:$NF/$(NF-1);++n} END{print x/n}' $log)
-	steal_avgok=$(awk '{x+=$NF==0?0:$(NF-4)/$NF;++n} END{print x/n}' $log)
+	steal_avgok=$(awk '{x+=$NF==0?0:$(NF-5)/$NF;++n} END{print x/n}' $log)
 	printf '%-25s\t%-9g\t%-9g\t%-9g\t%-9g\t%-9g\t%-9g\t%g\n'	\
 	    $label							\
 	    $mean $stddev $maxdev					\
diff --git a/libworkstream_df/test-cdeque.c b/libworkstream_df/test-cdeque.c
@@ -45,6 +45,7 @@ static void *worker_main (void *);
 static void worker (struct state *, unsigned long);
 static void straight_worker (struct state *, unsigned long);
 static void *thief_main (void *);
+static bool do_steal (struct state *, double *);
 
 static void *
 worker_main (void *data)
@@ -119,39 +120,51 @@ static void *
 thief_main (void *data)
 {
   struct state *state = data;
-  void *val;
-  double stealprob;
-  unsigned long i;
+  double lasttime;
 
   while (!atomic_load_explicit (&start, memory_order_acquire))
     continue;
 
   BEGIN_TIME (&state->time);
 
-  if (steal_freq > 0.0)
-    stealprob = steal_freq;
-  else
-    stealprob = (double) num_steal_per_thread / num_job;
-  for (i = 0; i < num_steal_per_thread; ++i)
+  lasttime = get_thread_cpu_time ();
+  while (state->num_attempt < num_steal_per_thread)
+    {
+      if (!do_steal (state, &lasttime))
+	break;
+    }
+
+  END_TIME (&state->time);
+
+  return NULL;
+}
+
+static bool
+do_steal (struct state *state, double *plasttime)
+{
+  double curtime;
+  unsigned long i, ndue;
+  void *val;
+
+  curtime = get_thread_cpu_time ();
+  ndue = (unsigned long) ((curtime - *plasttime) * steal_freq);
+  if (ndue == 0)
+    return !atomic_load_explicit (&end, memory_order_acquire);
+  *plasttime = curtime;
+  for (i = 0; i < ndue; ++i)
     {
-      /* Minimal probability is actually 1/RAND_MAX. */
-      while ((double) rand_r (&state->seed) / RAND_MAX > stealprob)
-	continue;
       val = cdeque_steal (worker_deque);
       if (val == NULL)
 	++state->num_failed_attempt;
       if (atomic_load_explicit (&end, memory_order_acquire))
 	{
 	  if (num_steal_per_thread == (unsigned long) -1)
-	    break;
+	    return false;
 	}
       else
 	++state->num_attempt;
     }
-
-  END_TIME (&state->time);
-
-  return NULL;
+  return true;
 }
 
 int
@@ -167,6 +180,7 @@ main (int argc, char *argv[])
   num_thread = 2;
   num_steal = -1;
   stealratio = -1.0;
+  steal_freq = 1000000.0;
   breadth = 6;
   depth = 10;
   dqlogsize = 6;
@@ -254,12 +268,10 @@ main (int argc, char *argv[])
       num_job += nrowjob;
     }
 
-  if (!has_num_steal && !has_steal_freq)
-    stealratio = -1.0;
   if (stealratio >= 0.0)
     num_steal = (unsigned long) (stealratio * num_job);
   if (num_steal == (unsigned long) -1)
-    num_steal_per_thread = -1;
+    num_steal_per_thread = (unsigned long) -1;
   else
     {
       num_steal_per_thread = num_steal / (num_thread - 1);
@@ -307,6 +319,7 @@ main (int argc, char *argv[])
   printf ("empty_takes\t"
 	  "empty_steals\t"
 	  "total_jobs\t"
+	  "steal_freq\t"
 	  "expected_steals\t"
 	  "effective_steals\n");
 
@@ -321,6 +334,7 @@ main (int argc, char *argv[])
   printf ("%-8lu\t", n);
 
   printf ("%-8lu\t", num_job);
+  printf ("%-8g\t", steal_freq);
   if (num_steal == (unsigned long) -1)
     printf ("unspecified\t");
   else
diff --git a/libworkstream_df/time-util.h b/libworkstream_df/time-util.h
@@ -29,6 +29,18 @@ do								\
   }								\
 while (0)
 
+static inline double
+get_thread_cpu_time (void)
+{
+  struct timespec _tv;
+  double _t;
+
+  clock_gettime (CLOCK_THREAD_CPUTIME_ID, &_tv);
+  _t = _tv.tv_sec;
+  _t += _tv.tv_nsec * 1e-9;
+  return _t;
+}
+
 static inline int
 timespec_diff (struct timespec *_result,
 	       const struct timespec *_px, const struct timespec *_py)
diff --git a/libworkstream_df/tune-test-cdeque.awk b/libworkstream_df/tune-test-cdeque.awk
