Grand Prix Garage - GPL

The workings of GPL Online - Part 2

Detailed notes on Bandwidth

From traces and replays I made using a server and one client in a local network, I noted the following figures. I checked them using an Internet connection between the systems as well. I simulated a multi client race by letting ai cars in.

• Upload packet size from client to server in all cases: 36 bytes, occasionally 16 bytes extra. The 84 bytes that are default are never fully used.
• Download packet size from server to client: net_xxx_server_send_size bytes.
  • The number of surrounding cars shown if the packet is completely filled is: (packet size - 4) / 16, so 5 cars for the default modem class packet size of 84.
  • If there are not enough cars to fill the maximum packet size, the packet size is 4 + 16 * [number of surrounding cars shown] + 16 bytes, with occasionally another 16 bytes extra. I don't know what the extra data of 16 or 32 bytes might be. So when hosting few clients on a modem connection, it may help to reduce mdm_server_send_size to 4+16*[number of surrounding cars shown], thereby preventing that the extra data is sent.
• For modem class connections, the default net_mdm_server_send_size of 84 is also the maximum. Specifying a greater value does not result in larger packets on the link. Therefor the maximum number of surrounding cars shown for a modem class connection is 5 (4 cars in front and 1 behind). The number can be reduced by specifying a smaller packet size.
• For lan class connections, there is no maximum. Since the maximum number of surrounding cars is 19, the largest packet that makes sense would be 4 + 32 + 16 * 19 = 340 bytes.
• The number of surrounding cars shown is split in cars ahead and cars behind as follows:
  • 1 - 3 cars shown: 1 -3 cars ahead, no cars behind (very dangerous unless only 1 - 3 other cars on track!)
  • 4 - 7 cars shown: 3 - 6 cars ahead, 1 car behind
  • 8 - 11 cars shown: 6 - 9 cars ahead, 2 cars behind
  • 12 - 15 cars shown: 9 - 12 cars ahead, 3 cars behind
  • 16 - 19 cars shown: 12 - 15 cars ahead, 4 cars behind (could not test >16 cars)
• The protocol overhead for all packets is 54 bytes:
  • UDP: 8 bytes
  • IP: 20 bytes
  • Link level:
    • Ethernet (cable & adsl connections): 26 bytes (8 bytes flag + 14 bytes header + 4 bytes CRC)
    • PPP (dial-up connections): 7 bytes (1 byte flag + 4 bytes header + 2 bytes FCS + 1 byte flag);
      extra bytes will be inserted into the data to maintain transparency. For reasons of simplicity I will use the Ethernet values for calculations as this is most common among hosts.
• To go from bytes to bits, I will multiply by 8 (based on Ethernet frames).
• TCP parameter settings play no role as all traffice is UDP. So there is no need to optimize MSS or DefaultRCVWindow. Since the largest IP packet GPL will send is 340 bytes application data + 28 bytes UDP/IP, there will be no segmentation even for the smallest value of MaxMTU (576).
• The client uses its local net_xxx_client_send_every setting to determine the send frequency of the local car data.
  The server uses its local net_xxx_server_send_every setting for the send frequency of the data on surrounding cars.
  The server uses its local net_xxx_server_send_size setting to determine the size of the packets it sends to the clients. If a client has a smaller value for net_xxx_server_send_size in its core.ini, a bandwidth mismatch will be reported when trying to connect.
  The net_xxx_client_send paramaters are ignored on the server.
  (xxx = lan when net_use_mdm_bandwidth_for_tcp_ip = 0
  xxx = mdm when net_use_mdm_bandwidth_for_tcp_ip = 1)
• The same amount of data is sent all the time, whether the local car is on the track or in the pits, in practise and in race.
  When the local car of the server never enters the track, its data is not sent to the clients.