For some reason std:: string/vector memory is much slower!

main.cpp

@@ -1,12 +1,25 @@
-#include "uWS.h"
+#include <string>
 
-std::string buffer;
+char largeBuf[] = "HTTP/1.1 200 OK\r\nContent-Length: 52428800\r\n\r\n";
+int largeHttpBufSize = sizeof(largeBuf) + 52428800 - 1;
+char *largeHttpBuf;
+
+#include "uWS.h"
 
 //#define USE_SSL
 
 int main() {
 
-    buffer = "Why hello there!";
+    std::string buffer;
+    buffer.resize(largeHttpBufSize);
+
+    std::vector<char> vec;
+    vec.resize(largeHttpBufSize);
+
+    // all STD-classes's memory cut throughput in half!
+    largeHttpBuf = (char *) buffer.data();//new char[largeHttpBufSize];//aligned_alloc(16, largeHttpBufSize);//std::aligned_alloc//vec.data();//malloc(largeHttpBufSize);
+    memcpy(largeHttpBuf, largeBuf, sizeof(largeBuf) - 1);
+    printf("%d\n", largeHttpBufSize);
 
     //uWS::init();
     //uWS::Loop loop();
@@ -22,15 +35,9 @@ int main() {
 
     app.onGet("/", [](auto *s, auto *req, auto *args) {
 
-        // this one is simple and okay for small sends
-        /*s->writeStatus("200 OK")
-         ->writeHeader("Server", "uWebSockets")
-         ->end(buffer);*/
-
-        // for large sends you want a stream!
-        s->writeStatus("200 OK")->write([](int offset) {
-            return std::string_view(buffer.data() + offset, buffer.length() - offset);
-        }, buffer.length());
+        s->/*writeStatus("200 OK")->*/write([](int offset) {
+            return std::string_view(largeHttpBuf + offset, largeHttpBufSize - offset);
+        }, largeHttpBufSize);
 
     }).onWebSocket("/wsApi", []() {
 

src/App.h

@@ -66,7 +66,7 @@ protected:
             static_dispatch(us_ssl_socket_context_on_open, us_socket_context_on_open)(httpServerContext, [](auto *s) {
                 Data *data = (Data *) static_dispatch(us_ssl_socket_context_ext, us_socket_context_ext)(static_dispatch(us_ssl_socket_get_context, us_socket_get_context)(s));
 
-                // here we need to construct a HTTP socket on the ext!
+                new (static_dispatch(us_ssl_socket_ext, us_socket_ext)(s)) typename HttpSocket<SSL>::Data;
 
                 if (!data->onHttpConnection) {
                     return;
@@ -78,6 +78,9 @@ protected:
             static_dispatch(us_ssl_socket_context_on_close, us_socket_context_on_close)(httpServerContext, [](auto *s) {
                 Data *data = (Data *) static_dispatch(us_ssl_socket_context_ext, us_socket_context_ext)(static_dispatch(us_ssl_socket_get_context, us_socket_get_context)(s));
 
+                // todo: run the destructor!
+                //((typename HttpSocket<SSL>::Data *) static_dispatch(us_ssl_socket_ext, us_socket_ext)(s))->~(typename HttpSocket<SSL>::Data)();
+
                 if (!data->onHttpDisconnection) {
                     return;
                 }
@@ -89,10 +92,6 @@ protected:
 
                 Data *contextData = (Data *) static_dispatch(us_ssl_socket_context_ext, us_socket_context_ext)(static_dispatch(us_ssl_socket_get_context, us_socket_get_context)(s));
 
-                    // a HttpSocket is basically the Http state and everything needed for it, we use it to parse the data and it knows about its context
-
-                    //
-
                 HttpRequest req(data, length);
                 if (req.isComplete()) {
                     contextData->onHttpRequest((HttpSocket<SSL> *) s, &req);
@@ -100,13 +99,17 @@ protected:
                     std::cout << "Got chunked HTTP headers!" << std::endl;
                 }
             });
+
+            static_dispatch(us_ssl_socket_context_on_writable, us_socket_context_on_writable)(httpServerContext, [](auto *s) {
+                ((HttpSocket<SSL> *) s)->onWritable();
+            });
         }
 
 public:
 
     // for server
     void listen(const char *host, int port, int options) {
-        static_dispatch(us_ssl_socket_context_listen, us_socket_context_listen)(httpServerContext, host, port, options, sizeof(HttpSocket<SSL>));
+        static_dispatch(us_ssl_socket_context_listen, us_socket_context_listen)(httpServerContext, host, port, options, sizeof(typename HttpSocket<SSL>::Data));
     }
 
     AppBase &onPost(std::string pattern, std::function<void(HttpSocket<SSL> *, HttpRequest *, std::vector<std::string_view> *)> handler) {

src/HttpSocket.h

@@ -10,6 +10,17 @@
 template <bool SSL>
 struct HttpSocket {
 
+    template <class A, class B>
+    static constexpr typename std::conditional<SSL, A, B>::type *static_dispatch(A *a, B *b) {
+        if constexpr(SSL) {
+            return a;
+        } else {
+            return b;
+        }
+    }
+
+    typedef typename std::conditional<SSL, us_ssl_socket, us_socket>::type SOCKET_TYPE;
+
     // chunked response will be tricky with this buffering scheme
     // if we do not fit, we can always use the header buffer for this (both in and out!)
     // put first 8kb chunk in the http buffer, then from there it's the stream's job!
@@ -21,7 +32,7 @@ struct HttpSocket {
         std::string headerBuffer;
         int offset = 0;
 
-        std::function<std::pair<const char *, int>(int)> outStream;
+        std::function<std::string_view(int)> outStream;
     };
 
     // only this one should be used!
@@ -51,13 +62,13 @@ struct HttpSocket {
     }
 
     // never rely on this one!
-    void writeToCorkBufferAndReset(const char *src, int length) {
+    void writeToCorkBufferAndReset(const char *src, int length, int contentLength) {
         uWS::Loop::Data *loopData = (uWS::Loop::Data *) us_loop_ext(us_socket_context_loop(us_socket_get_context((us_socket *) this)));
 
         memcpy(loopData->corkBuffer + loopData->corkOffset, "Content-Length: ", 16);
         loopData->corkOffset += 16;
 
-        loopData->corkOffset += u32toa_naive(length, loopData->corkBuffer + loopData->corkOffset);
+        loopData->corkOffset += u32toa_naive(contentLength, loopData->corkBuffer + loopData->corkOffset);
 
         memcpy(loopData->corkBuffer + loopData->corkOffset, "\r\n\r\n", 4);
         loopData->corkOffset += 4;
@@ -97,9 +108,42 @@ struct HttpSocket {
 
     // stream out (todo: fix up large sends and benchmark it again)
     void write(std::function<std::string_view(int)> cb, int length) {
+
+        // just assume this went fine
+        Data *httpData = (Data *) static_dispatch(us_ssl_socket_ext, us_socket_ext)((SOCKET_TYPE *) this);
+
+        //httpData->offset = 0;//uWS::Loop::MAX_COPY_DISTANCE;
+
+        // now we start streaming as much as possible in each call!
         std::string_view chunk = cb(0);
 
-        writeToCorkBufferAndReset(chunk.data(), chunk.length());
+        // write that off!
+        if constexpr (SSL) {
+
+        } else {
+            httpData->offset = us_socket_write((SOCKET_TYPE *) this, chunk.data(), chunk.length(), 0);
+        }
+
+        // if offset is at the end, we are done
+        if (httpData->offset < length) {
+            httpData->outStream = cb;
+        }
+    }
+
+    // this thing should only be reachable from App!
+    void onWritable() {
+        Data *httpData = (Data *) static_dispatch(us_ssl_socket_ext, us_socket_ext)((SOCKET_TYPE *) this);
+
+
+        // now we start streaming as much as possible in each call!
+        std::string_view chunk = httpData->outStream(httpData->offset);
+
+        // write that off!
+        if constexpr (SSL) {
+
+        } else {
+            httpData->offset += us_socket_write((SOCKET_TYPE *) this, chunk.data(), chunk.length(), 0);
+        }
     }
 
     // can't create this!

src/Loop.h

@@ -6,14 +6,18 @@
 #include <new>
 #include <string_view>
 #include <iostream>
+#include <unistd.h>
 
 namespace uWS {
 struct Loop {
     us_loop *loop;
 
+    static const int CORK_BUFFER_SIZE = 16 * 1024;
+    static const int MAX_COPY_DISTANCE = 4 * 1024;
+
     struct Data {
 
-        char *corkBuffer = new char[1024];
+        char *corkBuffer = new char[CORK_BUFFER_SIZE];
         int corkOffset = 0;
 
         Data() {