Merge pull request #298 from leondavi/results_opt

[ApiServer] Optimize communication of model phase results

Author: leondavi

src_cpp/opennnBridge/openNNnif.cpp

@@ -102,7 +102,8 @@ void* PredictFun(void* arg)
     // Stop the timer and calculate the time took for training
     high_resolution_clock::time_point  stop = high_resolution_clock::now();
     auto duration = duration_cast<microseconds>(stop - PredictNNptr->start_time);
-    nifpp::TERM predict_time = nifpp::make(env, duration.count());
+
+    ERL_NIF_TERM predict_time = enif_make_double(env, duration.count());
     nifpp::str_atom nerlnif_atom_str(NERLNIF_ATOM_STR);
     nifpp::TERM nerlnif_atom = nifpp::make(env , nerlnif_atom_str);
     ERL_NIF_TERM predict_res_and_time = enif_make_tuple(env, 4 , nerlnif_atom , prediction , nifpp::make(env, PredictNNptr->return_tensor_type) , predict_time);

src_erl/NerlnetApp/src/Bridge/nerlNIF.erl

@@ -53,8 +53,7 @@ call_to_train(ModelID, {DataTensor, Type}, WorkerPid , BatchID , SourceName)->
             {nerlnif, nan, TrainTime} -> 
                   gen_statem:cast(WorkerPid,{loss, nan , TrainTime , BatchID , SourceName}); %TODO Guy - Please the behavior when this case happens
             {nerlnif , LossTensor, LossTensorType , TrainTime}->
-                  {ErlTensor, ErlTensorType} = nerltensor_conversion({LossTensor, LossTensorType}, erl_float), % TODO Guy - Please do the conversion in main server
-                  gen_statem:cast(WorkerPid,{loss, {ErlTensor, ErlTensorType} , TrainTime , BatchID , SourceName})
+                  gen_statem:cast(WorkerPid,{loss, {LossTensor, LossTensorType} , TrainTime , BatchID , SourceName})
             after ?TRAIN_TIMEOUT ->  %TODO inspect this timeout 
                   ?LOG_ERROR("Worker train timeout reached! bid:~p s:~p",[BatchID , SourceName]),
                   gen_statem:cast(WorkerPid,{loss, timeout , SourceName}) %% TODO Guy Define train timeout state 
@@ -64,11 +63,11 @@ call_to_predict(ModelID, {BatchTensor, Type}, WorkerPid, BatchID , SourceName)->
       ok = predict_nif(ModelID, BatchTensor, Type),
       receive
 
-            {nerlnif , PredNerlTensor, NewType, TimeTook}-> %% nerlnif atom means a message from the nif implementation
+            {nerlnif , PredNerlTensor, PredNerlTensorType, TimeNif}-> %% nerlnif atom means a message from the nif implementation
                   % io:format("pred_nif done~n"),
                   % {PredTen, _NewType} = nerltensor_conversion({PredNerlTensor, NewType}, erl_float),
                   % io:format("Pred returned: ~p~n", [PredNerlTensor]),
-                  gen_statem:cast(WorkerPid,{predictRes,PredNerlTensor, NewType, TimeTook, BatchID , SourceName});
+                  gen_statem:cast(WorkerPid,{predictRes,PredNerlTensor, PredNerlTensorType, TimeNif, BatchID , SourceName});
             Error ->
                   ?LOG_ERROR("received wrong prediction_nif format: ~p" ,[Error]),
                   throw("received wrong prediction_nif format")

src_erl/NerlnetApp/src/Bridge/onnWorkers/workerGeneric.erl

@@ -161,24 +161,23 @@ idle(cast, _Param, State) ->
 
 %% Waiting for receiving results or loss function
 %% Got nan or inf from loss function - Error, loss function too big for double
-wait(cast, {loss , nan , TimeNIF , BatchID , SourceName}, State = #workerGeneric_state{myName = MyName, nextState = NextState}) ->
+wait(cast, {loss, nan , TrainTime , BatchID , SourceName}, State = #workerGeneric_state{myName = MyName, nextState = NextState}) ->
   stats:increment_by_value(get(worker_stats_ets), nan_loss_count, 1),
-  gen_statem:cast(get(client_pid),{loss, MyName , SourceName ,nan , TimeNIF ,BatchID}),
+  gen_statem:cast(get(client_pid),{loss, MyName , SourceName ,nan , TrainTime ,BatchID}),
   {next_state, NextState, State};
 
-wait(cast, {loss, LossTensor , TimeNIF , BatchID , SourceName}, State = #workerGeneric_state{myName = MyName, nextState = NextState, modelID=_ModelID, distributedBehaviorFunc = DistributedBehaviorFunc, distributedWorkerData = DistributedWorkerData}) ->
-  % {[_ , _ , _ , LossValue] , _} = LossTensor, 
-  % io:format("Got Loss Value ~p~n",[LossValue]),
+
+wait(cast, {loss, {LossTensor, LossTensorType} , TrainTime , BatchID , SourceName}, State = #workerGeneric_state{myName = MyName, nextState = NextState, modelID=_ModelID, distributedBehaviorFunc = DistributedBehaviorFunc, distributedWorkerData = DistributedWorkerData}) ->
   BatchTimeStamp = erlang:system_time(nanosecond),
-  gen_statem:cast(get(client_pid),{loss, MyName, SourceName ,LossTensor , TimeNIF , BatchID , BatchTimeStamp}), %% TODO Add Time and Time_NIF to the cast
+  gen_statem:cast(get(client_pid),{loss, MyName, SourceName ,{LossTensor, LossTensorType} , TrainTime , BatchID , BatchTimeStamp}),
   ToUpdate = DistributedBehaviorFunc(post_train, {get(generic_worker_ets),DistributedWorkerData}),
   if  ToUpdate -> {next_state, update, State#workerGeneric_state{nextState=NextState}};
       true ->     {next_state, NextState, State}
   end;
 
-wait(cast, {predictRes,PredNerlTensor, Type, TimeNIF, BatchID , SourceName}, State = #workerGeneric_state{myName = MyName, nextState = NextState, distributedBehaviorFunc = DistributedBehaviorFunc, distributedWorkerData = DistributedWorkerData}) ->
+wait(cast, {predictRes, PredNerlTensor, PredNerlTensorType, TimeNif, BatchID , SourceName}, State = #workerGeneric_state{myName = MyName, nextState = NextState, distributedBehaviorFunc = DistributedBehaviorFunc, distributedWorkerData = DistributedWorkerData}) ->
   BatchTimeStamp = erlang:system_time(nanosecond),
-  gen_statem:cast(get(client_pid),{predictRes,MyName,SourceName, {PredNerlTensor, Type}, TimeNIF , BatchID , BatchTimeStamp}), 
+  gen_statem:cast(get(client_pid),{predictRes,MyName, SourceName, {PredNerlTensor, PredNerlTensorType}, TimeNif , BatchID , BatchTimeStamp}), 
   Update = DistributedBehaviorFunc(post_predict, {get(generic_worker_ets),DistributedWorkerData}),
   if Update -> 
     {next_state, update, State#workerGeneric_state{nextState=NextState}};

src_erl/NerlnetApp/src/Client/clientStatem.erl

@@ -279,7 +279,7 @@ training(cast, _In = {predict}, State = #client_statem_state{myName = MyName, et
   ?LOG_ERROR("Wrong request , client ~p can't go from training to predict directly", [MyName]),
   {next_state, training, State#client_statem_state{etsRef = EtsRef}};
 
-training(cast, In = {loss , WorkerName , SourceName , LossTensor , TimeNIF , BatchID , BatchTS}, State = #client_statem_state{myName = MyName,etsRef = EtsRef}) ->
+training(cast, In = {loss, WorkerName ,SourceName ,LossTensor ,TimeNIF ,BatchID ,BatchTS}, State = #client_statem_state{myName = MyName,etsRef = EtsRef}) ->
   ClientStatsEts = get(client_stats_ets),
   stats:increment_messages_received(ClientStatsEts),
   stats:increment_bytes_received(ClientStatsEts , nerl_tools:calculate_size(In)),
@@ -315,14 +315,15 @@ predict(cast, In = {sample,Body}, State = #client_statem_state{etsRef = EtsRef})
   end,
   {next_state, predict, State#client_statem_state{etsRef = EtsRef}};
 
-predict(cast, In = {predictRes,WorkerName, SourceName ,{PredictNerlTensor, Type} , TimeTook , BatchID , BatchTS}, State = #client_statem_state{myName = _MyName, etsRef = EtsRef}) ->
+predict(cast, In = {predictRes,WorkerName, SourceName ,{PredictNerlTensor, NetlTensorType} , TimeTook , BatchID , BatchTS}, State = #client_statem_state{myName = _MyName, etsRef = EtsRef}) ->
   ClientStatsEts = get(client_stats_ets),
   stats:increment_messages_received(ClientStatsEts),
   stats:increment_bytes_received(ClientStatsEts , nerl_tools:calculate_size(In)),
- 
+
   {RouterHost,RouterPort} = ets:lookup_element(EtsRef, my_router, ?DATA_IDX),
-  MessageBody = {atom_to_list(WorkerName), SourceName, BatchID, {PredictNerlTensor , Type} , TimeTook , BatchTS}, %% SHOULD INCLUDE TYPE?
+  MessageBody = {WorkerName, SourceName, {PredictNerlTensor , NetlTensorType}, TimeTook, BatchID, BatchTS}, %% SHOULD INCLUDE TYPE?
   nerl_tools:http_router_request(RouterHost, RouterPort, [?MAIN_SERVER_ATOM], atom_to_list(predictRes), MessageBody),
+  
   stats:increment_messages_sent(ClientStatsEts),
   stats:increment_bytes_sent(ClientStatsEts , nerl_tools:calculate_size(MessageBody)),
   {next_state, predict, State#client_statem_state{etsRef = EtsRef}};

src_erl/NerlnetApp/src/MainServer/mainGenserver.erl

@@ -262,16 +262,16 @@ handle_cast({clientAck,Body}, State = #main_genserver_state{clientsWaitingList =
   ClientName = binary_to_term(Body),
   NewWaitingList = WaitingList--[ClientName], % waitingList is initialized in clientsTraining or clientsPredict handl cast calls
   if length(NewWaitingList) == 0 ->
-            ResultsToSendStr = generate_phase_result_data_to_send_from_ets_as_str(),
-            NothingToSend = string:is_empty(ResultsToSendStr),
+            PhaseResultsDataMap = generate_phase_result_data_map(),
+            NothingToSend = string:is_empty(PhaseResultsDataMap),
             if 
               NothingToSend -> pass;
               true ->  Action = case get(active_phase) of
                                         training -> trainRes;
                                         prediction -> predRes
                                     end,
                             {RouterHost,RouterPort} = ets:lookup_element(get(main_server_ets), my_router, ?DATA_IDX), % get main_server's router
-                            nerl_tools:http_router_request(RouterHost, RouterPort, [?API_SERVER_ATOM], atom_to_list(Action), ResultsToSendStr),
+                            nerl_tools:http_router_request(RouterHost, RouterPort, [?API_SERVER_ATOM], atom_to_list(Action), {json, PhaseResultsDataMap}),
                             stats:increment_messages_sent(StatsEts),
                             clean_phase_result_data_to_send_ets() % getting ready for next phase after data was sent to APIServer
             end,
@@ -324,11 +324,12 @@ handle_cast({lossFunction,Body}, State = #main_genserver_state{myName = MyName})
   stats:increment_messages_received(StatsEts),
   try
     case binary_to_term(Body) of
-        {WorkerName , SourceName , {LossTensor , _Type} , TimeNIF , BatchID , BatchTS} ->
-        ToSend = ?PHASE_RES_DATA_SEPARATOR ++ atom_to_list(WorkerName) ++ ?PHASE_RES_WORKER_NAME_SEPERATOR ++ atom_to_list(SourceName) ++ 
-                 ?PHASE_RES_VALUES_SEPERATOR ++ nerl_tools:string_format("~p",[LossTensor]) ++ ?PHASE_RES_VALUES_SEPERATOR ++ float_to_list(TimeNIF) ++ 
-                 ?PHASE_RES_VALUES_SEPERATOR ++ integer_to_list(BatchID) ++ ?PHASE_RES_VALUES_SEPERATOR ++ integer_to_list(BatchTS) ++ ?PHASE_RES_DATA_SEPARATOR,
-        store_phase_result_data_to_send_ets({WorkerName, BatchID , BatchTS}, ToSend);
+        {WorkerName , SourceName , {LossNerlTensor , LossNerlTensorType} , TimeNIF , BatchID , BatchTS} ->
+        Key = atom_to_list(WorkerName) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ atom_to_list(SourceName) ++ 
+              ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ integer_to_list(BatchID) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ 
+              integer_to_list(BatchTS) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ float_to_list(TimeNIF) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++
+              atom_to_list(LossNerlTensorType),
+        store_phase_result_data_to_send_ets(Key, binary_to_list(LossNerlTensor));
       _ELSE ->
         ?LOG_ERROR("~p Wrong loss function pattern received from client and its worker ~p", [MyName, Body])
     end
@@ -344,18 +345,12 @@ handle_cast({predictRes,Body}, State) ->
   _BatchSize = ets:lookup_element(get(main_server_ets), batch_size, ?DATA_IDX),
   stats:increment_messages_received(StatsEts),
   try 
-      {WorkerName, SourceName, BatchID, {NerlTensor, Type}, TimeNIF , BatchTS} = binary_to_term(Body),   %% TODO: add convention with client
-      %io:format("WorkerName: ~p, InputName: ~p, BatchID: ~p,  Type: ~p~n",[WorkerName, InputName, BatchID, Type]),
-      {DecodedNerlTensor, _Type} =
-      if 
-        (NerlTensor==<<>>) -> ?LOG_ERROR(?LOG_HEADER++"Got empty tensor"), empty_nerltensor_err;
-        true ->  nerlNIF:nerltensor_conversion({NerlTensor, Type}, nerlNIF:erl_type_conversion(Type)) % converting nerltensor from binary to erlang type using NerlNIF
-      end,
-      ToSend = ?PHASE_RES_DATA_SEPARATOR ++ WorkerName ++ ?PHASE_RES_WORKER_NAME_SEPERATOR ++ atom_to_list(SourceName) ++ 
-               ?PHASE_RES_VALUES_SEPERATOR ++ nerl_tools:string_format("~p",[DecodedNerlTensor]) ++ ?PHASE_RES_VALUES_SEPERATOR ++ 
-                integer_to_list(TimeNIF) ++ ?PHASE_RES_VALUES_SEPERATOR ++ integer_to_list(BatchID) ++ ?PHASE_RES_VALUES_SEPERATOR ++ 
-                integer_to_list(BatchTS) ++ ?PHASE_RES_DATA_SEPARATOR,
-      store_phase_result_data_to_send_ets({WorkerName, BatchID , BatchTS}, ToSend)
+      {WorkerName, SourceName, {NerlTensor, NerlTensorType}, TimeNIF , BatchID, BatchTS} = binary_to_term(Body),
+      Key = atom_to_list(WorkerName) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ atom_to_list(SourceName) ++ 
+            ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ integer_to_list(BatchID) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ 
+            integer_to_list(BatchTS) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++ float_to_list(TimeNIF) ++ ?PHASE_RES_VALUES_IN_KEY_SEPARATOR ++
+            atom_to_list(NerlTensorType),
+      store_phase_result_data_to_send_ets(Key, binary_to_list(NerlTensor))
   catch Err:E ->  
     ?LOG_ERROR(?LOG_HEADER++"Error receiving predict result ~p",[{Err,E}])
   end,
@@ -471,19 +466,14 @@ retransmission_to_apiserver(HttpRouterRequestFunc, Trials) ->
   end.
 
 
-store_phase_result_data_to_send_ets({WorkerName, BatchID , BatchTS}, DataToSendStr) -> 
-  Key = {WorkerName, BatchID , BatchTS},
-  ets:insert(get(phase_res_data_ets),{Key, DataToSendStr}).
-
+store_phase_result_data_to_send_ets(Key, NerlTensorData) ->
+  KeyBin = list_to_binary(Key),
+  ets:insert(get(phase_res_data_ets),{KeyBin, NerlTensorData}).
 
-generate_phase_result_data_string_from_list([], _ResString) -> _ResString;
-generate_phase_result_data_string_from_list(ListOfData, ResString) ->
-  NewResString = ResString++element(?DATA_IDX,hd(ListOfData)),
-  generate_phase_result_data_string_from_list(tl(ListOfData), NewResString).
 
-generate_phase_result_data_to_send_from_ets_as_str() ->
+generate_phase_result_data_map() ->
   ListOfData = ets:tab2list(get(phase_res_data_ets)),
-  generate_phase_result_data_string_from_list(ListOfData, ""). % String to send is retruned
+  ListOfData.
 
 clean_phase_result_data_to_send_ets() ->
   ets:delete_all_objects(get(phase_res_data_ets)).

src_erl/NerlnetApp/src/MainServer/mainServerDefs.hrl

@@ -1,5 +1,6 @@
 -define(API_SERVER_ACTION_ACK, "ackPy").
 
+-define(PHASE_RES_VALUES_IN_KEY_SEPARATOR, "#").
 -define(PHASE_RES_WORKER_NAME_SEPERATOR, "#").
 -define(PHASE_RES_VALUES_SEPERATOR, "|").
 -define(PHASE_RES_DATA_SEPARATOR, "?").

src_erl/NerlnetApp/src/Source/sourceStatem.erl

@@ -22,7 +22,7 @@
 
 
 %% defintions
--define(SENDING_FREQUENCY_OVERHEAD_FIX_FACTOR_PERC, 0.85).
+-define(SENDING_FREQUENCY_OVERHEAD_FIX_FACTOR_PERC, 0.75).
 -define(MICRO_TO_MILLI_FACTOR, 0.001).
 
 

src_erl/NerlnetApp/src/nerl_tools.erl

@@ -31,14 +31,22 @@ http_router_request(RouterHost, RouterPort, DestinationsList, ActionStr, Body) -
   end.
 
 
+http_request(Host, Port, Path, {json, Body}) -> 
+  io:format("Sending Json to ~p:~p~n",[Host,Port]),
+  JsonContentType = ?HTTP_CONTENT_TYPE_JSON,
+  Json = jsx:encode(Body),
+  http_request(Host, Port,Path, JsonContentType, Json);
+http_request(Host, Port,Path, Body) -> 
+  DefaultContentType = ?HTTP_CONTENT_TYPE_FORM_URLENCODED,
+  http_request(Host, Port,Path, DefaultContentType, Body).
 
 %% send message between entities
-http_request(Host, Port,Path, Body) when is_atom(Body) -> http_request(Host, Port,Path, atom_to_list(Body));
-http_request(Host, Port,Path, Body) when is_binary(Host) -> http_request(binary_to_list(Host), Port,Path, Body);
-http_request(Host, Port,Path, Body)->
-  URL = "http://" ++ Host ++ ":"++integer_to_list(Port) ++ "/" ++ Path,
+http_request(Host, Port, Path, ContentType, Body) when is_atom(Body) -> http_request(Host, Port,Path, ContentType, atom_to_list(Body));
+http_request(Host, Port, Path, ContentType, Body) when is_binary(Host) -> http_request(binary_to_list(Host), Port,Path, ContentType, Body);
+http_request(Host, Port, Path, ContentType, Body)->
+  URL = "http://" ++ Host ++ ":"++integer_to_list(Port) ++ "/" ++ Path, % Path is the action
   httpc:set_options([{proxy, {{Host, Port},[Host]}}]),
-  httpc:request(post,{URL, [],"application/x-www-form-urlencoded",Body}, [], []).
+  httpc:request(post,{URL, [], ContentType, Body}, [], []).
 
 get_client_worker_pairs([],_WorkersMap,Ret)-> Ret;
 get_client_worker_pairs([WorkerName|WorkersNames],WorkersMap,Ret)->

src_erl/NerlnetApp/src/nerl_tools.hrl

@@ -13,6 +13,11 @@
 -define(VALIDATION_OF_TRANSMISSION_WITH_API_SERVER_INTERVAL_MS, 100). % how much between each resend
 %% ETS definitions 
 
+%% HTTP Content type definitions
+-define(HTTP_CONTENT_TYPE_MULTI_PART_FORM_DATA, "multipart/form-data").
+-define(HTTP_CONTENT_TYPE_JSON, "application/json").
+-define(HTTP_CONTENT_TYPE_FORM_URLENCODED, "application/x-www-form-urlencoded").
+
 % 2 elements ETS:
 -define(KEY_IDX, 1).
 -define(DATA_IDX, 2).