diff --git a/LibCarla/source/carla/rpc/Server.h b/LibCarla/source/carla/rpc/Server.h
index 6b51e7c7c..8f647eccc 100644
--- a/LibCarla/source/carla/rpc/Server.h
+++ b/LibCarla/source/carla/rpc/Server.h
@@ -51,6 +51,7 @@ namespace rpc {
     }
 
     void SyncRunFor(time_duration duration) {
+      TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
       _sync_io_context.reset();
       _sync_io_context.run_for(duration.to_chrono());
     }
diff --git a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Game/CarlaEngine.cpp b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Game/CarlaEngine.cpp
index 96a5285ec..1b393421c 100644
--- a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Game/CarlaEngine.cpp
+++ b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Game/CarlaEngine.cpp
@@ -111,6 +111,7 @@ void FCarlaEngine::NotifyEndEpisode()
 
 void FCarlaEngine::OnPreTick(UWorld *, ELevelTick TickType, float DeltaSeconds)
 {
+  TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
   if (TickType == ELevelTick::LEVELTICK_All)
   {
     // update frame counter
@@ -132,6 +133,7 @@ void FCarlaEngine::OnPreTick(UWorld *, ELevelTick TickType, float DeltaSeconds)
 
 void FCarlaEngine::OnPostTick(UWorld *World, ELevelTick TickType, float DeltaSeconds)
 {
+  TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
   // tick the recorder/replayer system
   if (GetCurrentEpisode())
   {
diff --git a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.cpp b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.cpp
index 742d2e899..405e9462d 100644
--- a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.cpp
+++ b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.cpp
@@ -46,6 +46,7 @@ static void WritePixelsToBuffer_Vulkan(
     uint32 Offset,
     FRHICommandListImmediate &InRHICmdList)
 {
+  TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
   check(IsInRenderingThread());
   auto RenderResource =
       static_cast<const FTextureRenderTarget2DResource *>(RenderTarget.Resource);
@@ -58,12 +59,18 @@ static void WritePixelsToBuffer_Vulkan(
   FIntPoint Rect = RenderResource->GetSizeXY();
 
   // NS: Extra copy here, don't know how to avoid it.
-  InRHICmdList.ReadSurfaceData(
-      Texture,
-      FIntRect(0, 0, Rect.X, Rect.Y),
-      gPixels,
-      FReadSurfaceDataFlags(RCM_UNorm, CubeFace_MAX));
-  Buffer.copy_from(Offset, gPixels);
+  {
+    TRACE_CPUPROFILER_EVENT_SCOPE_TEXT("Read Surface");
+    InRHICmdList.ReadSurfaceData(
+        Texture,
+        FIntRect(0, 0, Rect.X, Rect.Y),
+        gPixels,
+        FReadSurfaceDataFlags(RCM_UNorm, CubeFace_MAX));
+  }
+  {
+    TRACE_CPUPROFILER_EVENT_SCOPE_TEXT("Buffer Copy");
+    Buffer.copy_from(Offset, gPixels);
+  }
 }
 
 // =============================================================================
@@ -130,6 +137,7 @@ void FPixelReader::WritePixelsToBuffer(
     FRHICommandListImmediate &InRHICmdList
     )
 {
+  TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
   check(IsInRenderingThread());
 
   if (IsVulkanPlatform(GMaxRHIShaderPlatform) || IsD3DPlatform(GMaxRHIShaderPlatform, false))
diff --git a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.h b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.h
index 00d8cf743..17d57d351 100644
--- a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.h
+++ b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/PixelReader.h
@@ -101,6 +101,8 @@ void FPixelReader::SendPixelsInRenderThread(TSensor &Sensor)
   (
     [&Sensor, Stream=Sensor.GetDataStream(Sensor)](auto &InRHICmdList) mutable
     {
+      TRACE_CPUPROFILER_EVENT_SCOPE_TEXT("FWritePixels_SendPixelsInRenderThread");
+
       /// @todo Can we make sure the sensor is not going to be destroyed?
       if (!Sensor.IsPendingKill())
       {
@@ -113,12 +115,16 @@ void FPixelReader::SendPixelsInRenderThread(TSensor &Sensor)
         if(Buffer.data())
         {
           SCOPE_CYCLE_COUNTER(STAT_CarlaSensorStreamSend);
+          TRACE_CPUPROFILER_EVENT_SCOPE_TEXT("Stream Send");
           Stream.Send(Sensor, std::move(Buffer));
         }
       }
     }
   );
 
-  // Blocks until the render thread has finished all it's tasks
-  Sensor.WaitForRenderThreadToFinsih();
+  {
+    TRACE_CPUPROFILER_EVENT_SCOPE_TEXT("FPixelReader::WaitForRenderThreadToFinsih");
+    // Blocks until the render thread has finished all it's tasks
+    Sensor.WaitForRenderThreadToFinsih();
+  }
 }
diff --git a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/WorldObserver.cpp b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/WorldObserver.cpp
index 2225f796e..5f8f0f177 100644
--- a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/WorldObserver.cpp
+++ b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Sensor/WorldObserver.cpp
@@ -179,7 +179,7 @@ static carla::Buffer FWorldObserver_Serialize(
     bool MapChange,
     bool PendingLightUpdates)
 {
-
+  TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
   using Serializer = carla::sensor::s11n::EpisodeStateSerializer;
   using SimulationState = carla::sensor::s11n::EpisodeStateSerializer::SimulationState;
   using ActorDynamicState = carla::sensor::data::ActorDynamicState;
@@ -243,6 +243,7 @@ void FWorldObserver::BroadcastTick(
     bool MapChange,
     bool PendingLightUpdates)
 {
+  TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
   auto AsyncStream = Stream.MakeAsyncDataStream(*this, Episode.GetElapsedGameTime());
 
   auto buffer = FWorldObserver_Serialize(
diff --git a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Server/CarlaServer.cpp b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Server/CarlaServer.cpp
index a2f963b6e..597d24a46 100644
--- a/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Server/CarlaServer.cpp
+++ b/Unreal/CarlaUE4/Plugins/Carla/Source/Carla/Server/CarlaServer.cpp
@@ -1664,14 +1664,31 @@ void FCarlaServer::AsyncRun(uint32 NumberOfWorkerThreads)
 {
   check(Pimpl != nullptr);
   /// @todo Define better the number of threads each server gets.
-  auto RPCThreads = NumberOfWorkerThreads / 2u;
-  auto StreamingThreads = NumberOfWorkerThreads - RPCThreads;
-  Pimpl->Server.AsyncRun(std::max(2u, RPCThreads));
-  Pimpl->StreamingServer.AsyncRun(std::max(2u, StreamingThreads));
+  int32_t RPCThreads = std::max(2u, NumberOfWorkerThreads / 2u);
+  int32_t StreamingThreads = std::max(2u, NumberOfWorkerThreads - RPCThreads);
+
+  UE_LOG(LogCarla, Error, TEXT("FCommandLine %s"), FCommandLine::Get());
+
+  if(!FParse::Value(FCommandLine::Get(), TEXT("-RPCThreads="), RPCThreads))
+  {
+    RPCThreads = std::max(2u, NumberOfWorkerThreads / 2u);
+  }
+  if(!FParse::Value(FCommandLine::Get(), TEXT("-StreamingThreads="), StreamingThreads))
+  {
+    StreamingThreads = std::max(2u, NumberOfWorkerThreads - RPCThreads);
+  }
+
+  UE_LOG(LogCarla, Error, TEXT("FCarlaServer AsyncRun %d, RPCThreads %d, StreamingThreads %d"),
+        NumberOfWorkerThreads, RPCThreads, StreamingThreads);
+
+  Pimpl->Server.AsyncRun(RPCThreads);
+  Pimpl->StreamingServer.AsyncRun(StreamingThreads);
+
 }
 
 void FCarlaServer::RunSome(uint32 Milliseconds)
 {
+  TRACE_CPUPROFILER_EVENT_SCOPE_TEXT(__FUNCTION__);
   Pimpl->Server.SyncRunFor(carla::time_duration::milliseconds(Milliseconds));
 }