Age of Empires
As with Tribes, the real-time strategy (RTS) game Age of Empires was released in the late 1990s. This means that Age of Empires faced many of the same bandwidth and latency constraints of dial-up Internet access. Age of Empires uses a deterministic lockstep networking model. In this model, all the computers are connected to each other, meaning it is peer-to-peer. A guaranteed deterministic simulation of the game is concurrently performed by each of the peers. It is lockstep because peers use communication to ensure that they remain synchronized throughout the game. As with Tribes, even though the deterministic lockstep model has existed for many years, it is still commonly used in modern RTS games. The other sample game built during the course of this book, RoboCat RTS, implements a deterministic lockstep model.
One of the largest differences between implementing networked multiplayer for an RTS instead of an FPS is the number of relevant units. In Tribes, even though there are up to 128 players, at any particular point in time only a fraction of these players is going to be relevant to a particular client. This means that the ghost manager in Tribes rarely has to send information about more than 20 to 30 ghosts at a time.
Contrast this with an RTS such as Age of Empires. Although the player cap is much smaller (limited to eight simultaneous players in the original game), each player can control a large number of units. The original Age of Empires capped the number of units for each player at 50, whereas in later games the cap was as high as 200. Using the cap of 50, this means that in a massive eight-player battle, there could be up to 400 units active at a time. Although it is natural to wonder if some sort of relevancy system could reduce the number of units that need to be synchronized, it’s important to consider the worst-case scenario. What if a battle toward the end of a game featured the armies of all eight players? In this case, there are going to be several hundred units that are relevant at the same time. It would be hard for the synchronization to keep up even if a minimal amount of information is sent per unit.
To alleviate this issue, the engineers for Age of Empires decided to synchronize the commands each player issued, rather than synchronizing the units. There’s a subtle but important distinction in this implementation—even a professional RTS player may be able to issue no more than 300 commands per minute. This means that even in an extreme case, the game need only transmit a few commands per second per each player. This requires a much more manageable amount of bandwidth than transmitting information about several hundred units. However, given that the game is no longer transmitting unit information over the network, each instance of the game needs to independently apply the commands transmitted by each player. Since each game instance is performing an independent simulation, it is of the utmost importance that each game instance remains synchronized with the other game instances. This ends up being the largest challenge of implementing the deterministic lockstep model.
Since every game instance is performing an independent simulation, it makes sense to utilize a peer-to-peer topology. As discussed in Chapter 6, “Network Topologies and Sample Games,” one advantage of a peer-to-peer model is that data can reach every computer more quickly. This is because the server is not acting as a middleman. However, one disadvantage is that each player needs to send their information to every other player, as opposed to just a single server. So for example, if player A issues an attack command, then every game instance needs to be aware of this attack command, or their simulations would diverge from each other.
However, there is another key factor to consider. Different players are going to run the game at different frame rates, and different players are going to have different quality connections. Going back to the example where player A issues an attack command, it’s just as important that player A does not immediately apply the attack command. Instead, player A should only apply the attack command once players B, C, and D are all ready to simultaneously apply the command. But this introduces a conundrum: If player A’s game waits too long to execute the attack command, the game will seem very unresponsive.
The solution to this problem is to introduce a turn timer to queue up commands. With the turn timer approach, first a turn length is selected—in the case of Age of Empires, the default duration was 200 ms. All commands during these 200 ms are saved into a buffer. When the 200 ms are over, all the commands for that player’s turn are transmitted over the network to all other players. Another key aspect of this system is a turn execution delay of two turns. What this means is that, for example, commands that are issued by the player on turn 50 will not be executed by any game until turn 52. In the case of a 200-ms turn timer, this means that the input lag, the amount of time it takes for a player’s command to be displayed on screen, could be as high as 600 ms. However, the two turns of slack allows for every other player to receive and acknowledge the commands for a particular turn. It may seem slightly counterintuitive for an RTS game to actually have turns, but you can see the hallmarks of the turn timer approach in many different RTS games, including StarCraft II. Of course, modern games can have the luxury of shorter turn timers since bandwidth and latency are much better for most users today in comparison to the late 1990s.
There is one important edge case to consider with the turn timer approach. What happens if one of the players experiences a lag spike and they can no longer keep up with the 200-ms timer? Some games might temporarily pause the simulation to see if the lag spike can be overcome—eventually, the game may decide to drop the player if they continue to slow down the game for everyone else. Age of Empires also tries to compensate for this scenario by dynamically adjusting the rendering frame rate based on network conditions—thus a computer with a particularly slow Internet connection might allocate more time to receive data over the network, with less time being allotted for rendering graphics. For more detail on the dynamic turn adjustment, consult the original Bettner and Terrano article listed in the references.
There’s also an extra benefit of transmitting the commands issued by the clients. With such an approach, it does not take much extra memory or work to save the commands issued over the course of an entire match. This directly leads to the possibility of implementing savable match replays, as in Age of Empires II. Replays are very popular in RTS games because it allows players to evaluate matches to gain a deeper understanding of strategies. It would require significantly more memory and overhead to create replays in an approach that transmitted unit information instead of commands.
Turn timers alone are not enough to guarantee synchronization between each peer. Since each machine is receiving and processing commands independently, it is of the utmost importance that each machine arrives at an identical result. In their paper, Bettner and Terrano write that “the difficulty with finding out-of-sync errors is that very subtle differences would multiply over time. A deer slightly out of alignment when the random map was created would forage slightly differently—and minutes later a villager would path a tiny bit off, or miss with his spear and take home no meat.”
One concrete example arises from the fact that most games have some amount of randomness in actions. For instance, what if the game performs a random check in order to determine whether or not an archer hits an infantry? It would be conceivable that player A’s instance decides the archer does hit the infantry, whereas player B’s instance decides the archer doesn’t hit the infantry. The solution to this problem is to exploit the “pseudo” prefix of the pseudo-random number generator (PRNG). Since all PRNGs use some sort of seeding, the way you can guarantee both players A and B arrive at the same random results is to synchronize the seed value across all game instances. One should keep in mind, however, that a seed only guarantees a particular sequence of numbers. So not only is it important that each game instance uses the same seed, it’s equally important that each game instance makes the same number of calls to the random generation number—otherwise the PRNG numbers will become out of sync. PRNG synchronization in a peer-to-peer configuration is further elaborated in Chapter 6, “Network Topologies and Sample Games.”
There is also an implicit advantage to checking for synchronization—it reduces the opportunity for players to cheat. For example, if one player gives themselves 500 extra resources, the other game instances could immediately detect the desynchronization in the game state. It would then be trivial to kick the offending player out of the game. However, as with any system, there are tradeoffs—the fact that each game state simulates each unit in the game means that it is possible to create cheats that reveal information that should not be visible. This means that the so-called “map hacks” that reveal the entire map are still a common issue in most RTS games. This and other security concerns are covered in Chapter 10, “Security.”