Latency of Reliable Streams

2020-04-10

This is the second article in a series in which I investigate what the ideal network protocol for games looks like. You can find the last article here.

Unreliable sequenced delivery is the workhorse of most online multiplayer game netcode, but with the diversity of netcode and game loops in the world I think trying to define, measure, and optimize a general solution to this is a fool’s errand. There are a few parameters to improve this delivery method for packet loss, such as pacing of packet transmission and backpressure on the netcode based on bandwidth estimations. Exposing these parameters to users is sufficient.

I was not excited to reach this boring conclusion, but fortunately there remains an interesting area to measure and benchmark: low latency reliable ordered streams of periodic payloads. These are employed by a few genres of discrete step games, games using state synchronization served by distributed systems, and many games that really ought to be using something else.

Contenders

I chose three protocols to measure: TCP as a reference, ENet as the status quo, and KCP as the state of the art.

I chose ENet solely based on its unrivaled popularity. I chose KCP because it claims to be state of the art and has significant adoption.

QUIC is excluded because it does not claim to be designed for this use case and measuring protocols takes work.

Limitations of Comparison

Comparing protocols is complicated because protocols are complicated.

A route between two sockets, which the protocols compared in the following benchmarks build on top of, is stateful and opaque. At any given time it has a certain amount of bandwidth available, a certain latency to deliver the packets, and probabilities of re-ordering and dropping packets. None of these attributes are simple to measure, independent of each other, or independent of the behavior of the protocol attempting to adapt to them.

For a fun example, what do you think will happen to a TCP connection that is sending 400 byte payloads at 60 Hertz on a connection with a router in the middle that has an outgoing token buffer filter rate limit of 200kbps (not enough) but no MTU and an unlimited queue size? The graph below illustrates the round trip time in milliseconds for these periodic payloads against the index of the payload.

tcp growth

As TCP sends more data than the rate limit will allow to pass without delay, the payload takes longer to arrive and ACK, but the payload does arrive and doesn’t drop. TCP adjusts to the delay by sending a larger packet. Because of the rate limit, the larger packet takes even longer to arrive. This repeats indefinitely; packets endlessly grow in size and take longer to arrive. The last packet sent on this chart was 32,834 bytes.

This combination of variables and many others that are easy to imagine will simply break protocols. The scenario demonstrated here obviously does not occur in real life, but a protocol designer has to know whether a scenario will occur, and how often, in many non-obvious cases.

Protocol designers must not only identify and accommodate most realistic network conditions, but also transitions between them. Even if the tech along the route behaves deterministically, the other users sharing the route will not. There will be bursts of packet loss, packet reordering, and router queuing delay.

On the other side, a protocol user’s send patterns can be completely inappropriate for the network conditions and the protocol designer must decide how to moderate their behavior.

The point I make here is that the performance of a protocol is a function of the conditions, and those are as diverse as real numbers. Any meaningful comparison of protocols is hyper specific and predictive only in similar conditions.

Benchmark

The benchmark is run in two network conditions:

The benchmark client periodically sends payloads of 400 bytes at 60 Hertz to a server. The server simply sends the payloads back. We measure the round trip time.

See the full datasheet and benchmark code.

In the simulation of normal conditions, both KCP and ENet maintain lower and less variant latency than TCP.

ENet consistently holds lower and less variant latency than KCP. ENet’s mean latency over 10 runs is 26.297ms (deviation 9.693ms) where KCP Turbo’s mean latency over 10 runs is 37.074ms (deviation 11.218ms). ENet round trips the payload in 71% of the time it takes KCP to round trip the same payload.

In the turbulent conditions, only KCP Turbo avoids significant latency spikes. KCP Turbo holds significantly lower latency than ENet, with a mean round trip time of 40.582ms (deviation 10.399ms). ENet’s mean round trip time was 139.306ms (deviation 147.850ms). KCP Turbo round trips the payload in 29% of the time it takes ENet to round trip the same payload.

Conclusion

ENet’s reliable ordered streams achieve lower and less variant latency for periodic payloads than KCP on unaltered connections I have observed in the United States. Under high packet loss, KCP performs better than ENet by an order.

KCP regresses to a fair proportion of bandwidth at a lower rate than ENet or KCP, backing off at 1.5x rather than 2x, and it aggressively proactively retransmits messages. If in the future KCP came to represent some meaningful proportion of traffic, I wonder if these results would hold. I can’t imagine it is in routers’ interests to reward this behavior, but I know little about fairness enforcement.

A combined solution is interesting to consider: a user of ENet could respond to congestion by reducing their send rate, or doubling down and communicating over KCP. KCP holds much lower latency than ENet under high packet loss, but at 10% loss it is still almost double what ENet gets without packet loss. It's possible certain use cases could benefit from switching to KCP instead of respecting the congestion.