Merge pull request #101 from georges-hatem/main

ghatem: first multi-threaded attempt

Author: gcarreno

entries/ghatem-fpc/README.md

@@ -132,4 +132,54 @@ Instead of extracting the station name as a string 1B times, and use it as a dic
 This requires us to migrate from TFPHashList to the generic TDictionary. Even though TDictionary is slower than TFPHashList, the overall improvements yielded a significant performance gain.
 Using TDictionary, the code is more similar to my Delphi version, in case we get to run those tests on a Windows PC.
 
-** expected timing: ~60 seconds, single-threaded**
+* expected timing: ~60 seconds, single-threaded*
+** ACTUAL TIMING: 58 seconds as per gcarreno **
+
+
+## Multi-Threaded Attempt (2024-04-10)
+
+In a first attempt, using 2 threads, I evaluated how harmful were my "shared-memory" variables, namely:
+ - the pre-allocated records array
+ - its counter
+ - the dictionary of aggregated data
+ - the stringlist of station names
+
+Using a Critical-Section, even on just the stringlist, the performance drops quite a bit.
+
+### minimize shared-memory
+
+I got rid of the pre-allocated records array and its counter, back to on-the-fly allocation of records
+Got rid of the station names stringlist, because merging those lists at the end was very slow: instead, store the station name directly in the record.
+Replaced the singular dictionary with an array of dictionaries, 1-per-thread.
+
+### redundant work
+
+All of this causes some redundant work to be made by various threads:
+ - one record is allocated per-station-per-thread
+ - since we store the station name in the record, we convert it to string (using `SetString`) once per-station-per-thread
+
+### parallelism
+
+Given N bytes and K threads, a basic attempt is to distribute a range `N / K` of data per thread.
+A thread may get its start/end boundaries in the middle of a line: we ensure that each line is processed exactly once.
+
+Again, just to evaluate how things will go, we wait for all threads to complete before proceeding.
+
+### merging
+
+Given 2 dictionaries L and R, we merge them into L. This merge is applied K-1 times, and dictionaries[0] will contain the aggregated result.
+We could later consider merging every 2 threads that have completed their work.
+
+### expectations
+
+parallelization on my environment (see above) performed quite poorly, considering there is no shared-mem to protect against concurrent access:
+4 threads were barely getting 60% improvement over 1 thread.
+
+is it due to:
+ - VM virtualization?
+ - too many dictionaries were causing too many cache misses?
+ - work-load distributed unevenly?
+
+Better wait and see the results on the real environment, before judging.
+
+

entries/ghatem-fpc/src/OneBRCproj.lpr

@@ -16,6 +16,8 @@
 
   TOneBRCApp = class(TCustomApplication)
   private
+    FFileName: string;
+    procedure RunOneBRC;
   protected
     procedure DoRun; override;
   public
@@ -26,13 +28,65 @@   TOneBRCApp = class(TCustomApplication)
 
 { TOneBRCApp }
 
-procedure TOneBRCApp.DoRun;
+procedure TOneBRCApp.RunOneBRC;
 var
-  ErrorMsg: String;
   vOneBRC: TOneBRC;
-  vFileName: string;
   vStart: Int64;
   vTime: Int64;
+begin
+  vOneBRC := TOneBRC.Create (32);
+  try
+    try
+      vOneBRC.mORMotMMF(FFileName);
+      vOneBRC.DispatchThreads;
+      vOneBRC.WaitAll;
+      vOneBRC.MergeAll;
+      vOneBRC.GenerateOutput;
+
+      //vStart := GetTickCount;
+      //vOneBRC.mORMotMMF (FFileName);
+      //vTime := GetTickCount - vStart;
+      //WriteLn('read: ' + FloatToStr(vTime / 1000));
+      //WriteLn('-----------');
+      //WriteLn;
+      //
+      //vStart := GetTickCount;
+      //vOneBRC.DispatchThreads;
+      //vOneBRC.WaitAll;
+      //vTime := GetTickCount - vStart;
+      //WriteLn('process: ' + FloatToStr(vTime / 1000));
+      //WriteLn('-----------');
+      //WriteLn;
+      //
+      //vStart := GetTickCount;
+      //vOneBRC.MergeAll;
+      //vTime := GetTickCount - vStart;
+      //WriteLn('merge: ' + FloatToStr(vTime / 1000));
+      //WriteLn('-----------');
+      //WriteLn;
+      //
+      //vStart := GetTickCount;
+      //vOneBRC.GenerateOutput;
+      //vTime := GetTickCount - vStart;
+      //WriteLn('generate: ' + FloatToStr(vTime / 1000));
+      //WriteLn('-----------');
+      //WriteLn;
+      //ReadLn;
+    except
+      on E: Exception do
+      begin
+        WriteLn(Format(rsErrorMessage, [ E.Message ]));
+        ReadLn;
+      end;
+    end;
+  finally
+    vOneBRC.Free;
+  end;
+end;
+
+procedure TOneBRCApp.DoRun;
+var
+  ErrorMsg: String;
 begin
   // quick check parameters
   ErrorMsg:= CheckOptions(Format('%s%s%s:',[
@@ -67,7 +121,7 @@ procedure TOneBRCApp.DoRun;
   end;
 
   if HasOption(cShortOptInput, cLongOptInput) then begin
-    vFileName := GetOptionValue(
+    FFileName := GetOptionValue(
       cShortOptInput,
       cLongOptInput
     );
@@ -78,45 +132,9 @@ procedure TOneBRCApp.DoRun;
     Exit;
   end;
 
-  vFileName := ExpandFileName(vFileName);
+  FFileName := ExpandFileName(FFileName);
 
-  vOneBRC := TOneBRC.Create;
-  try
-    try
-      vOneBRC.mORMotMMF(vFileName);
-      vOneBRC.SingleThread;
-      vOneBRC.GenerateOutput;
-
-      {vStart := GetTickCount;
-      vOneBRC.mORMotMMF (vFileName);
-      vTime := GetTickCount - vStart;
-      WriteLn('read: ' + FloatToStr(vTime / 1000));
-      WriteLn('-----------');
-      WriteLn;
-
-      vStart := GetTickCount;
-      vOneBRC.SingleThread;
-      vTime := GetTickCount - vStart;
-      WriteLn('process: ' + FloatToStr(vTime / 1000));
-      WriteLn('-----------');
-      WriteLn;
-
-      vStart := GetTickCount;
-      vOneBRC.GenerateOutput;
-      vTime := GetTickCount - vStart;
-      WriteLn('generate: ' + FloatToStr(vTime / 1000));
-      WriteLn('-----------');
-      WriteLn;
-      ReadLn; }
-    except
-      on E: Exception do
-      begin
-        WriteLn(Format(rsErrorMessage, [ E.Message ]));
-      end;
-    end;
-  finally
-    vOneBRC.Free;
-  end;
+  RunOneBRC;
 
   // stop program loop
   Terminate;