Improve benchmark

1 files changed, 205 insertions, 19 deletions

diff --git a/lib/ssl/test/ssl_dist_bench_SUITE.erl b/lib/ssl/test/ssl_dist_bench_SUITE.erl
index 3c7904cf24..7409b69639 100644
--- a/lib/ssl/test/ssl_dist_bench_SUITE.erl
+++ b/lib/ssl/test/ssl_dist_bench_SUITE.erl
@@ -32,6 +32,8 @@
 -export(
    [setup/1,
     roundtrip/1,
+    throughput_0/1,
+    throughput_64/1,
     throughput_1024/1,
     throughput_4096/1,
     throughput_16384/1,
@@ -55,7 +57,9 @@ groups() ->
      {setup, [{repeat, 1}], [setup]},
      {roundtrip, [{repeat, 1}], [roundtrip]},
      {throughput, [{repeat, 1}],
-      [throughput_1024,
+      [throughput_0,
+       throughput_64,
+       throughput_1024,
        throughput_4096,
        throughput_16384,
        throughput_65536,
@@ -247,8 +251,9 @@ setup(A, B, Prefix, HA, HB) ->
     [] = ssl_apply(HB, erlang, nodes, []),
     {SetupTime, CycleTime} =
         ssl_apply(HA, fun () -> setup_runner(A, B, Rounds) end),
-    [] = ssl_apply(HA, erlang, nodes, []),
-    [] = ssl_apply(HB, erlang, nodes, []),
+    ok = ssl_apply(HB, fun () -> setup_wait_nodedown(A, 10000) end),
+    %% [] = ssl_apply(HA, erlang, nodes, []),
+    %% [] = ssl_apply(HB, erlang, nodes, []),
     SetupSpeed = round((Rounds*1000000*1000) / SetupTime),
     CycleSpeed = round((Rounds*1000000*1000) / CycleTime),
     _ = report(Prefix++" Setup", SetupSpeed, "setups/1000s"),
@@ -275,6 +280,22 @@ setup_loop(A, B, T, N) ->
             setup_loop(A, B, Time + T, N - 1)
     end.
 
+setup_wait_nodedown(A, Time) ->
+    ok = net_kernel:monitor_nodes(true),
+    case nodes() of
+        [] ->
+            ok;
+        [A] ->
+            receive
+                {nodedown,A} ->
+                    ok;
+                Unexpected ->
+                    {error,{unexpected,Unexpected}}
+            after Time ->
+                    {error,timeout}
+            end
+    end.
+
 
 %%----------------
 %% Roundtrip speed
@@ -334,6 +355,18 @@ roundtrip_client(Pid, Mon, StartTime, N) ->
 %%-----------------
 %% Throughput speed
 
+throughput_0(Config) ->
+    run_nodepair_test(
+      fun (A, B, Prefix, HA, HB) ->
+              throughput(A, B, Prefix, HA, HB, 500000, 0)
+      end, Config).
+
+throughput_64(Config) ->
+    run_nodepair_test(
+      fun (A, B, Prefix, HA, HB) ->
+              throughput(A, B, Prefix, HA, HB, 500000, 64)
+      end, Config).
+
 throughput_1024(Config) ->
     run_nodepair_test(
       fun (A, B, Prefix, HA, HB) ->
@@ -373,45 +406,198 @@ throughput_1048576(Config) ->
 throughput(A, B, Prefix, HA, HB, Packets, Size) ->
     [] = ssl_apply(HA, erlang, nodes, []),
     [] = ssl_apply(HB, erlang, nodes, []),
-    Time =
+    #{time := Time,
+      dist_stats := DistStats,
+      client_msacc_stats := ClientMsaccStats,
+      client_prof := ClientProf,
+      server_msacc_stats := ServerMsaccStats,
+      server_prof := ServerProf} =
         ssl_apply(HA, fun () -> throughput_runner(A, B, Packets, Size) end),
     [B] = ssl_apply(HA, erlang, nodes, []),
     [A] = ssl_apply(HB, erlang, nodes, []),
-    Speed = round((Packets*Size*1000000) / (1024*Time)),
+    ClientMsaccStats =:= undefined orelse
+        msacc:print(ClientMsaccStats),
+    io:format("DistStats: ~p~n", [DistStats]),
+    Overhead =
+        50 % Distribution protocol headers (empirical) (TLS+=54)
+        + byte_size(erlang:term_to_binary([0|<<>>])), % Benchmark overhead
+    Bytes = Packets * (Size + Overhead),
+    io:format("~w bytes, ~.4g s~n", [Bytes,Time/1000000]),
+    ClientMsaccStats =:= undefined orelse
+        io:format(
+          "Sender core usage ratio: ~.4g ns/byte~n",
+          [msacc:stats(system_runtime, ClientMsaccStats)*1000/Bytes]),
+    ServerMsaccStats =:= undefined orelse
+        begin
+            io:format(
+              "Receiver core usage ratio: ~.4g ns/byte~n",
+              [msacc:stats(system_runtime, ServerMsaccStats)*1000/Bytes]),
+            msacc:print(ServerMsaccStats)
+        end,
+    io:format("******* ClientProf:~n", []), prof_print(ClientProf),
+    io:format("******* ServerProf:~n", []), prof_print(ServerProf),
+    Speed = round((Bytes * 1000000) / (1024 * Time)),
     report(Prefix++" Throughput_"++integer_to_list(Size), Speed, "kB/s").
 
 %% Runs on node A and spawns a server on node B
 throughput_runner(A, B, Rounds, Size) ->
     Payload = payload(Size),
-    ClientPid = self(),
     [A] = rpc:call(B, erlang, nodes, []),
+    ClientPid = self(),
     ServerPid =
         erlang:spawn(
           B,
           fun () -> throughput_server(ClientPid, Rounds) end),
     ServerMon = erlang:monitor(process, ServerPid),
-    microseconds(
-      throughput_client(
-        ServerPid, ServerMon, Payload, start_time(), Rounds)).
+    msacc:available() andalso
+        begin
+            msacc:stop(),
+            msacc:reset(),
+            msacc:start(),
+            ok
+        end,
+    prof_start(),
+    {Time,ServerMsaccStats,ServerProf} =
+        throughput_client(ServerPid, ServerMon, Payload, Rounds),
+    prof_stop(),
+    ClientMsaccStats =
+        case msacc:available() of
+            true ->
+                MStats = msacc:stats(),
+                msacc:stop(),
+                MStats;
+            false ->
+                undefined
+        end,
+    ClientProf = prof_end(),
+    [{_Node,Socket}] = dig_dist_node_sockets(),
+    DistStats = inet:getstat(Socket),
+    #{time => microseconds(Time),
+      dist_stats => DistStats,
+      client_msacc_stats => ClientMsaccStats,
+      client_prof => ClientProf,
+      server_msacc_stats => ServerMsaccStats,
+      server_prof => ServerProf}.
+
+dig_dist_node_sockets() ->
+    [case DistCtrl of
+         {_Node,Socket} = NodeSocket when is_port(Socket) ->
+             NodeSocket;
+         {Node,DistCtrlPid} when is_pid(DistCtrlPid) ->
+             [{links,DistCtrlLinks}] = process_info(DistCtrlPid, [links]),
+             case [S || S <- DistCtrlLinks, is_port(S)] of
+                 [Socket] ->
+                     {Node,Socket};
+                 [] ->
+                     [{monitors,[{process,DistSenderPid}]}] =
+                         process_info(DistCtrlPid, [monitors]),
+                     [{links,DistSenderLinks}] =
+                         process_info(DistSenderPid, [links]),
+                     [Socket] = [S || S <- DistSenderLinks, is_port(S)],
+                     {Node,Socket}
+             end
+     end || DistCtrl <- erlang:system_info(dist_ctrl)].
+
 
-throughput_server(_Pid, 0) ->
-    ok;
 throughput_server(Pid, N) ->
+    msacc:available() andalso
+        begin
+            msacc:stop(),
+            msacc:reset(),
+            msacc:start(),
+            ok
+        end,
+    prof_start(),
+    throughput_server_loop(Pid, N).
+
+throughput_server_loop(_Pid, 0) ->
+    prof_stop(),
+    MsaccStats =
+        case msacc:available() of
+            true ->
+                msacc:stop(),
+                MStats = msacc:stats(),
+                msacc:reset(),
+                MStats;
+            false ->
+                undefined
+        end,
+    Prof = prof_end(),
+    exit({ok,MsaccStats,Prof});
+throughput_server_loop(Pid, N) ->
     receive
-        [N|_] ->
-            throughput_server(Pid, N-1)
+        {Pid, N, _} ->
+            throughput_server_loop(Pid, N-1)
     end.
 
-throughput_client(_Pid, Mon, _Payload, StartTime, 0) ->
+throughput_client(Pid, Mon, Payload, N) ->
+    throughput_client_loop(Pid, Mon, Payload, N, start_time()).
+
+throughput_client_loop(_Pid, Mon, _Payload, 0, StartTime) ->
     receive
-        {'DOWN', Mon, _, _, normal} ->
-            elapsed_time(StartTime);
+        {'DOWN', Mon, _, _, {ok,MsaccStats,Prof}} ->
+            {elapsed_time(StartTime),MsaccStats,Prof};
         {'DOWN', Mon, _, _, Other} ->
             exit(Other)
     end;
-throughput_client(Pid, Mon, Payload, StartTime, N) ->
-    Pid ! [N|Payload],
-    throughput_client(Pid, Mon, Payload, StartTime, N - 1).
+throughput_client_loop(Pid, Mon, Payload, N, StartTime) ->
+    Pid ! {self(), N, Payload},
+    throughput_client_loop(Pid, Mon, Payload, N - 1, StartTime).
+
+
+-define(prof, none). % none | cprof | eprof
+
+-if(?prof =:= cprof).
+prof_start() ->
+    cprof:stop(),
+    cprof:start(),
+    ok.
+-elif(?prof =:= eprof).
+prof_start() ->
+    {ok,_} = eprof:start(),
+    profiling = eprof:start_profiling(processes()),
+    ok.
+-elif(?prof =:= none).
+prof_start() ->
+    ok.
+-endif.
+
+-if(?prof =:= cprof).
+prof_stop() ->
+    cprof:pause(),
+    ok.
+-elif(?prof =:= eprof).
+prof_stop() ->
+    _ = eprof:stop_profiling(),
+    ok.
+-elif(?prof =:= none).
+prof_stop() ->
+    ok.
+-endif.
+
+-if(?prof =:= cprof).
+prof_end() ->
+    Prof = cprof:analyse(),
+    cprof:stop(),
+    Prof.
+-elif(?prof =:= eprof).
+prof_end() ->
+    eprof:dump_data().
+-elif(?prof =:= none).
+prof_end() ->
+    [].
+-endif.
+
+-if(?prof =:= cprof).
+prof_print(Prof) ->
+    io:format("~p.~n", [Prof]).
+-elif(?prof =:= eprof).
+prof_print(Dump) ->
+    eprof:analyze(undefined, total, [], Dump).
+-elif(?prof =:= none).
+prof_print([]) ->
+    ok.
+-endif.
 
 %%%-------------------------------------------------------------------
 %%% Test cases helpers