Counterstrike application

Counterstrike is a team-based multi-player game where online players are grouped into competing teams where each team is assigned a specific goal. The environment of the game is pre-loaded and clients exchange game state over the network using small UDP packets. Bursty losses can have an adverse effect on the progress of this game. First, during the initiation phase, the client generates important control packets which if lost can render the client unable to connect to the server. Second, a burst of packet losses during the middle of a game can either cause skips or cause a player to get disconnected. A skip can arise because the game state messages received immediately after a congestion provides a context jump in the game. Third, in a multi-player game, problems observed by one player will affect other players in the game. For example, disconnection of a single player can sometimes halt the progress of a game.

OverQoS can alleviate the problem of bursty losses by performing the following operations:

1. Recover from bursty network losses by using an FEC+ARQ based CLVL abstraction between overlay links along the path.
2. Smoothly drop data packets equivalent to the size of the burst at the overlay node.
3. Identify control packets based on packet size and not drop these packets.

=By both recovering lost packets as well as smoothly dropping an equivalent amount of data packets at an overlay node, OverQoS achieves three objectives: (a) OverQoS provides the Counterstrike client with critical updates to continue the progress of the game. For example, many UDP packets generated by Counterstrike merely contain the coordinates of different players. If OverQoS can deliver even a fraction of these packets, the client will still be able to reconstruct the movement and position of other players. (b) OverQoS uses minimal amount of additional bandwidth to support this application. The additional bandwidth which is not compensated by OverQoS is the FEC portion of the redundant traffic. Our wide-area experiments over realistic overlay links show that this additional bandwidth is negligible (refer to Section 6). (c) The application observes the same loss-rate as it would in the normal Internet yet not experience any skips in a game. In the event of a bursty loss, the application experiences an additional delay equal to the loss recovery time of a CLVL. With a reasonable distribution of overlay nodes, we expect this recovery time to be much smaller than end-to-end recovery.

Counterstrike Proxy: By reverse engineering the traffic characteristics of Counterstrike, we built a client and server proxy to interpret the Counterstrike packets. We chose a proxy-based implementation for two reasons: First, Counterstrike client and server codes are proprietary and we do not modify the code. Second, it is a simple way of capturing different application specific traffic and tunneling them through OverQoS.

Figure 5: Snapshot from a Counterstrike game at a $10\%$ loss rate.

Figure 6: Sequence number plot illustrating smoothing of packet losses using OverQoS.

Example Scenario: We consider a cable modem loss trace with an high loss-rate of $10\%$ and compare the effect of losses on the Counterstrike game under two scenarios: (a) with OverQoS; (b) without OverQoS. Figure 5 illustrates a snapshot of a Counterstrike game where OverQoS converts bursty losses into smooth losses and the client does not observe any skips. Figure 6 better illustrates the smoothing of losses using OverQoS. In the case without OverQoS, we observed many short periods of time where the network losses was as high as $70-80\%$ followed by periods with no congestion. The OverQoS node compensates the additional bandwidth consumed for loss recovery by smoothly dropping packets during non-lossy periods.

We make two additional observations. First, smoothing losses works well only when the bursty loss-periods are relatively short by compensating. When burst periods last for longer periods of time, OverQoS will not be able to smoothly drop packets in the absence of any non-lossy periods. Second, in this scenario, the CLVL abstraction is unable to achieve the target loss-rate due to congestion periods with very high loss-rates. However, the loss reduction provided by OverQoS during bursty periods is sufficient for the Counterstrike game to progress.