Unix域套接字非常慢

Question

我得到的代码（ 客户端 ； 服务器 ）需要在进程之间非常快速地传输数据（以仿真可扩展的内存总线；有效地，我尝试将Gameboy仿真器的“ Game Cartette Emulation”部分与GameGame仿真器的“ Game System Emulation”部分分开，以便可以轻松实现/测试/调试/其他更多虚拟和真实的硬件）。

为此，我选择了使用SOCK_SEQPACKET unix域套接字。

没有错误处理的简化示例：

typedef struct {
  enum {
    CONNECT,
    CONNECT_RESP,
    //... a few more packet types
  } type;
  uint16_t addr;
  uint16_t len; //note this is not the packet length, but a length field for a specific operation omitted for simplicity in the type field above
  uint8_t value;
} packet;

int server_main(int argc, char *argv[]) {
  const char* socket_path= "\0de.nonchip.PlutoBoy.hook_mem";
  int fd = socket(AF_UNIX, SOCK_SEQPACKET, 0);
  struct sockaddr_un addr;
  memset(&addr, 0, sizeof(addr));
  addr.sun_family = AF_UNIX;
  *addr.sun_path = '\0';
  strncpy(addr.sun_path+1, socket_path+1, sizeof(addr.sun_path)-2);
  size_t psize=sizeof(packet);
  setsockopt(fd, SOL_SOCKET, SO_SNDBUF, &psize, sizeof(psize));
  bind(fd, (struct sockaddr*)&addr, sizeof(addr));
  listen(fd, 5);
  packet p;
  while(1){
    int cl = accept(fd, NULL, NULL));
    while(recv(cl,&p,psize,MSG_WAITALL)>0){
      if(p.type==CONNECT)
        p.type=CONNECT_RESP;
      send(cl,&p,psize,MSG_EOR);
    }
  }
}

//----

void client_connect_socket(){
  const char* socket_path = "\0de.nonchip.PlutoBoy.hook_mem";
  struct sockaddr_un socket_addr;
  int socket_fd = socket(AF_UNIX, SOCK_SEQPACKET, 0)
  memset(&socket_addr, 0, sizeof(socket_addr));
  socket_addr.sun_family = AF_UNIX;
  *socket_addr.sun_path = '\0';
  strncpy(socket_addr.sun_path+1, socket_path+1, sizeof(socket_addr.sun_path)-2);

  setsockopt(socket_fd, SOL_SOCKET, SO_SNDBUF, &psize, sizeof(psize));
  connect(socket_fd, (struct sockaddr*)&socket_addr, sizeof(socket_addr));

  packet p;
  p.type=CONNECT;
  send(socket_fd, &p, psize,MSG_EOR);
  recv(socket_fd, &p, psize,MSG_WAITALL);
  if(p.type != CONNECT_RESP){
    perror("connect handshake error");
    exit(-1);
  }
}

但是这样做确实很慢，并且将系统负载提高到正常值的3-4倍，所以我决定为其添加原始基准测试（使用CLOCK_MONOTONIC和字节计数器），这给了我：

  211091 bytes transferred in  5 s =    42218 B/s
  213748 bytes transferred in  5 s =    42749 B/s
  213062 bytes transferred in  5 s =    42612 B/s
  215458 bytes transferred in  5 s =    43091 B/s
  219934 bytes transferred in  5 s =    43986 B/s
  210272 bytes transferred in  5 s =    42054 B/s
  216723 bytes transferred in  5 s =    43344 B/s

抢先内核上的两个进程之间通过unix域套接字的连接如何只有42KB / s？ 这真的很正常吗？还是我在这里错过了一些重要的东西？

编辑：我创建了一些客户端和服务器的 callgrind日志，但看不到任何东西真正在那儿阻塞了我。