Merge branch 'lukas/ssl/dist_benchmarks'

* lukas/ssl/dist_benchmarks:
  stdlib: Adjust binary bench to be faster
  ssl: Add scheduler utilization benchmark of plain/tls dist
  ssl: Link to tstsrvr to group_leader in test
  ssl: Add some benchmarks to test

4 files changed, 123 insertions, 15 deletions

diff --git a/lib/ssl/test/ssl.spec b/lib/ssl/test/ssl.spec
index cb54168d36..5e65bfcfe6 100644
--- a/lib/ssl/test/ssl.spec
+++ b/lib/ssl/test/ssl.spec
@@ -1,4 +1,9 @@
-{suites,"../ssl_test",all}.
-{skip_suites, "../ssl_test",
-     [ssl_bench_SUITE, ssl_dist_bench_SUITE],
-     "Benchmarks run separately"}.
+% {merge_tests,false}.
+{alias,dir,"../ssl_test"}.
+
+{suites,dir,all}.
+{skip_groups,dir,ssl_bench_SUITE,setup,"Benchmarks run separately"}.
+{skip_groups,dir,ssl_bench_SUITE,pem_cache,"Benchmarks run separately"}.
+{skip_groups,dir,ssl_dist_bench_SUITE,setup,"Benchmarks run separately"}.
+{skip_groups,dir,ssl_dist_bench_SUITE,throughput,"Benchmarks run separately"}.
+
diff --git a/lib/ssl/test/ssl_dist_bench_SUITE.erl b/lib/ssl/test/ssl_dist_bench_SUITE.erl
index 3c7904cf24..daedd254d7 100644
--- a/lib/ssl/test/ssl_dist_bench_SUITE.erl
+++ b/lib/ssl/test/ssl_dist_bench_SUITE.erl
@@ -32,6 +32,7 @@
 -export(
    [setup/1,
     roundtrip/1,
+    sched_utilization/1,
     throughput_1024/1,
     throughput_4096/1,
     throughput_16384/1,
@@ -54,6 +55,7 @@ groups() ->
      %%
      {setup, [{repeat, 1}], [setup]},
      {roundtrip, [{repeat, 1}], [roundtrip]},
+     {sched_utilization,[{repeat, 1}], [sched_utilization]},
      {throughput, [{repeat, 1}],
       [throughput_1024,
        throughput_4096,
@@ -65,7 +67,9 @@ groups() ->
 all_groups() ->
     [{group, setup},
      {group, roundtrip},
-     {group, throughput}].
+     {group, throughput},
+     {group, sched_utilization}
+    ].
 
 init_per_suite(Config) ->
     Digest = sha1,
@@ -330,6 +334,98 @@ roundtrip_client(Pid, Mon, StartTime, N) ->
             roundtrip_client(Pid, Mon, StartTime, N - 1)
     end.
 
+%%---------------------------------------
+%% Scheduler utilization at constant load
+
+
+sched_utilization(Config) ->
+    run_nodepair_test(
+      fun(A, B, Prefix, HA, HB) ->
+              sched_utilization(A, B, Prefix, HA, HB, proplists:get_value(ssl_dist, Config))
+      end, Config).
+
+sched_utilization(A, B, Prefix, HA, HB, SSL) ->
+    [] = ssl_apply(HA, erlang, nodes, []),
+    [] = ssl_apply(HB, erlang, nodes, []),
+    {ClientMsacc, ServerMsacc, Msgs} =
+        ssl_apply(HA, fun () -> sched_util_runner(A, B, SSL) end),
+    [B] = ssl_apply(HA, erlang, nodes, []),
+    [A] = ssl_apply(HB, erlang, nodes, []),
+    msacc:print(ClientMsacc),
+    msacc:print(ServerMsacc),
+    ct:pal("Got ~p msgs",[length(Msgs)]),
+    report(Prefix++" Sched Utilization Client",
+           10000 * msacc:stats(system_runtime,ClientMsacc) /
+               msacc:stats(system_realtime,ClientMsacc), "util 0.01 %"),
+    report(Prefix++" Sched Utilization Server",
+           10000 * msacc:stats(system_runtime,ServerMsacc) /
+               msacc:stats(system_realtime,ServerMsacc), "util 0.01 %"),
+    ok.
+
+%% Runs on node A and spawns a server on node B
+%% We want to avoid getting busy_dist_port as it hides the true SU usage
+%% of the receiver and sender.
+sched_util_runner(A, B, true) ->
+    sched_util_runner(A, B, 50);
+sched_util_runner(A, B, false) ->
+    sched_util_runner(A, B, 250);
+sched_util_runner(A, B, Senders) ->
+    Payload = payload(5),
+    [A] = rpc:call(B, erlang, nodes, []),
+    ServerPid =
+        erlang:spawn(
+          B,
+          fun () -> throughput_server() end),
+    ServerMsacc =
+        erlang:spawn(
+          B,
+          fun() ->
+                  receive
+                      {start,Pid} ->
+                          msacc:start(10000),
+                          Pid ! {ServerPid,msacc:stats()}
+                  end
+          end),
+    ServerMon = erlang:monitor(process, ServerPid),
+    ClientPid =
+        spawn_link(fun() ->
+                           %% We spawn 250 senders which should mean that we
+                           %% have a load of 250 msgs/msec
+                           [spawn_link(fun() -> throughput_client(ServerPid,Payload) end) ||
+                               _ <- lists:seq(1, Senders)]
+                   end),
+
+    erlang:system_monitor(self(),[busy_dist_port]),
+    ServerMsacc ! {start,self()},
+    msacc:start(10000),
+    ClientMsaccStats = msacc:stats(),
+    ServerMsaccStats = receive {ServerPid,Stats} -> Stats end,
+    {ClientMsaccStats,ServerMsaccStats, flush()}.
+
+flush() ->
+    receive
+        M ->
+            [M | flush()]
+    after 0 ->
+            []
+    end.
+
+throughput_server() ->
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    receive _ -> ok end,
+    throughput_server().
+
+throughput_client(Pid, Payload) ->
+    Pid ! Payload,
+    receive after 1 -> throughput_client(Pid, Payload) end.
 
 %%-----------------
 %% Throughput speed
diff --git a/lib/ssl/test/ssl_dist_test_lib.erl b/lib/ssl/test/ssl_dist_test_lib.erl
index 1b9c853fc4..b8b805b2c4 100644
--- a/lib/ssl/test/ssl_dist_test_lib.erl
+++ b/lib/ssl/test/ssl_dist_test_lib.erl
@@ -144,6 +144,8 @@ mk_node_cmdline(ListenPort, Name, Args) ->
 	++ integer_to_list(ListenPort) ++ " "
 	++ Args ++ " "
 	++ "-env ERL_CRASH_DUMP " ++ Pwd ++ "/erl_crash_dump." ++ Name ++ " "
+        ++ "-kernel inet_dist_connect_options \"[{recbuf,12582912},{sndbuf,12582912},{high_watermark,8388608},{low_watermark,4194304}]\" "
+        ++ "-kernel inet_dist_listen_options \"[{recbuf,12582912},{sndbuf,12582912},{high_watermark,8388608},{low_watermark,4194304}]\" "
 	++ "-kernel error_logger \"{file,\\\"" ++ Pwd ++ "/error_log." ++ Name ++ "\\\"}\" "
 	++ "-setcookie " ++ atom_to_list(erlang:get_cookie()).
 
@@ -179,8 +181,9 @@ check_ssl_node_up(Socket, Name, Bin) ->
 		    Parent = self(),
 		    Go = make_ref(),
 		    %% Spawn connection handler on test server side
-		    Pid = spawn_link(
+		    Pid = spawn(
 			    fun () ->
+                                    link(group_leader()),
 				    receive Go -> ok end,
                                     process_flag(trap_exit, true),
 				    tstsrvr_con_loop(Name, Socket, Parent)
diff --git a/lib/stdlib/test/stdlib_bench_SUITE.erl b/lib/stdlib/test/stdlib_bench_SUITE.erl
index b937eeb06a..436aa78a0f 100644
--- a/lib/stdlib/test/stdlib_bench_SUITE.erl
+++ b/lib/stdlib/test/stdlib_bench_SUITE.erl
@@ -38,7 +38,9 @@ groups() ->
       [norm_nfc_list, norm_nfc_deep_l, norm_nfc_binary,
        string_lexemes_list, string_lexemes_binary
       ]},
-     {binary, [{repeat, 5}],
+     %% Only run 1 binary match repeat as it is very slow pre OTP-22.
+     %% The results seem to be stable enough anyway
+     {binary, [{repeat, 1}],
       [match_single_pattern_no_match,
        matches_single_pattern_no_match,
        matches_single_pattern_eventual_match,
@@ -164,19 +166,19 @@ norm_data(Config) ->
 
 match_single_pattern_no_match(_Config) ->
     Binary = binary:copy(<<"ugbcfuysabfuqyfikgfsdalpaskfhgjsdgfjwsalp">>, 1000000),
-    comment(test(binary, match, [Binary, <<"o">>])).
+    comment(test(100, binary, match, [Binary, <<"o">>])).
 
 matches_single_pattern_no_match(_Config) ->
     Binary = binary:copy(<<"ugbcfuysabfuqyfikgfsdalpaskfhgjsdgfjwsalp">>, 1000000),
-    comment(test(binary, matches, [Binary, <<"o">>])).
+    comment(test(100, binary, matches, [Binary, <<"o">>])).
 
 matches_single_pattern_eventual_match(_Config) ->
     Binary = binary:copy(<<"ugbcfuysabfuqyfikgfsdalpaskfhgjsdgfjwsal\n">>, 1000000),
-    comment(test(binary, matches, [Binary, <<"\n">>])).
+    comment(test(100, binary, matches, [Binary, <<"\n">>])).
 
 matches_single_pattern_frequent_match(_Config) ->
     Binary = binary:copy(<<"abc\n">>, 1000000),
-    comment(test(binary, matches, [Binary, <<"abc">>])).
+    comment(test(100, binary, matches, [Binary, <<"abc">>])).
 
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 
@@ -233,9 +235,11 @@ list() ->
     random_byte_list(?SIZE).
 
 test(Mod, Fun, Args) ->
-    F = fun() -> loop(?N, Mod, Fun, Args) end,
+    test(?N, Mod, Fun, Args).
+test(Iter, Mod, Fun, Args) ->
+    F = fun() -> loop(Iter, Mod, Fun, Args) end,
     {Time, ok} = timer:tc(fun() -> lspawn(F) end),
-    report_mfa(Time, Mod).
+    report_mfa(Iter, Time, Mod).
 
 loop(0, _M, _F, _A) -> garbage_collect(), ok;
 loop(N, M, F, A) ->
@@ -248,8 +252,8 @@ lspawn(Fun) ->
         {'DOWN', Ref, process, Pid, Rep} -> Rep
     end.
 
-report_mfa(Time, Mod) ->
-    Tps = round((?N*1000000)/Time),
+report_mfa(Iter, Time, Mod) ->
+    Tps = round((Iter*1000000)/Time),
     ct_event:notify(#event{name = benchmark_data,
                            data = [{suite, "stdlib_" ++ atom_to_list(Mod)},
                                    {value, Tps}]}),