Hi, I’m Glenn Fiedler and welcome to the first article in my article series Networking for Game Programmers

In this article we start with the most basic aspect of network programming, sending and receiving data over the network. This is just the beginning – the simplest and most basic part of what network programmers do, but still it is quite intricate and non-obvious as to what the best course of action is. Take care because if you get this part wrong it will have terrible effects on your multiplayer game!

You have most likely heard of sockets, and are probably aware that there are two main types: TCP and UDP. When writing a network game, we first need to choose what type of socket to use. Do we use TCP sockets, UDP sockets or a mixture of both?

The choice you make depends entirely on what sort of game you want to network. So from this point on, and for the rest of this article series, I’m going to assume you want to network an action game. You know games like Halo, Battlefield 1942, Quake, Unreal, CounterStrke, Team Fortress and so on.

In light of the fact that we want to network an action game, we’ll take a very close look at the properties of each type of socket, and dig a bit into how the internet actually works. Once we have all this information, the correct choice becomes very clear!


TCP stands for “transmission control protocol” and IP stands for “internet protocol”. Together they form the backbone for almost everything you do online, from web browsing to IRC to email, its all built on top of TCP/IP.

If you have ever used a TCP socket, then you’ll know that it is a reliable connection based protocol. This simply means that you make a connection between two machines, then you send data between the two computers much like you are writing to a file on one side, and reading from a file on the other.

This connection is reliable and ordered, meaning that all data you send is guaranteed to arrive at the other side in the same order that you wrote it. Its also a stream of data, meaning that TCP takes care of splitting up your data into packets and sending those across the network for you.

Again, remember its just like writing to a file. So simple!


The simplicity is in stark contrast to what actually goes on at the lower level “IP” protocol underneath TCP.

Here there is no concept of connection, instead packets are passed from one computer to the next. You can visualize this process like a hand-written note being passed from one person to the next across a crowded room, eventually reaching the person it is addressed to, but only after passing through many hands.

There is no guarantee that this note will actually reach the person it is addressed to. The sender just passes the note along and hopes for the best, never knowing whether or not the note was received, unless the other person decides to write back!

Of course, it is in reality a little bit more complicated than this, since of course no one computer knows the exact sequence of computers to pass the packet along to so that it reaches its destination quickly. Sometimes “IP” passes along multiple copies of the same packet, these packets make their way to the destination via different paths, so will most likely arrive at different times.

This is because the internet is designed to be self-organizing and self-repairing, able to route around connectivity problems. It’s actually quite cool if you think about what it is really going on at the low level. You can read all about this in the classic book TCP/IP Illustrated.


Instead of treating communications between computers like writing to files, what if we want to send and receive packets directly?

We can do this using UDP. UDP stands for “user datagram protocol” and it is another protocol built on top of IP, just like TCP, but this time instead of adding lots of features and complexity it is just a very thin layer over IP.

With UDP we can send a packet to a destination IP address (eg. and port (say 52423), and it will get passed from computer to computer until it arrives at the destination computer or is lost along the way.

On the receiver side, we just sit there listening on a specific port (eg. 52423) and when a packet arrives from any computer (remember there are no connections!), we get notified of the address and port of the computer that sent the packet, the size of the packet, and can read the packet data.

UDP is an unreliable protocol. In practice, most packets that are sent will get through, but you’ll usually have around 1-5% packet loss, and occasionally you’ll get periods where no packets get through at all (remember that there are lots of computers between you and your destination where things can go wrong…)

There is also no guarantee of ordering of packets. You could send 5 packets in order 1,2,3,4,5 and they could arrive completely out of order like 3,1,2,5,4. In practice, they will actually arrive in order almost all of the time, but again, you cannot rely on this!

Finally, although UDP doesn’t do much on top of IP, it does make one guarantee for you. If you send a packet, it will either arrive in whole at the destination, or not arrive at all. So if you send a 256 byte packet to another computer, that computer cannot receive just the first 100 bytes of the packet, it must get the full 256 bytes of data. This is pretty much the only guarantee you get with UDP, everything else is up to you!


We have a decision to make here, do we use TCP sockets or UDP sockets?

Lets look at the properties of each:


  • Connection based
  • Guaranteed reliable and ordered
  • Automatically breaks up your data into packets for you
  • Makes sure it doesn’t send data too fast for the internet connection to handle (flow control)
  • Easy to use, you just read and write data like its a file


  • No concept of connection, you have to code this yourself
  • No guarantee of reliability or ordering of packets, they may arrive out of order, be duplicated, or not arrive at all!
  • You have to manually break your data up into packets and send them
  • You have to make sure you don’t send data too fast for your internet connection to handle
  • If a packet is lost, you need to devise some way to detect this, and resend that data if necessary

The decision seems pretty clear then, TCP does everything we want and its super easy to use, while UDP is a huge pain in the ass and we have to code everything ourselves from scratch. So obviously we just use TCP right?


Using TCP is the worst possible mistake you can make when developing a networked action game like an FPS! To understand why, you need to see what TCP is actually doing above IP to make everything look so simple!

How TCP really works

TCP and UDP are both built on top of IP, but they are radically different. UDP behaves very much like the IP protocol underneath it, while TCP abstracts everything so it looks like you are reading and writing to a file, hiding all complexities of packets and unreliability from you.

So how does it do this?

Firstly, TCP is a stream protocol, so you just write bytes to a stream, and TCP makes sure that they get across to the other side. Since IP is built on packets, and TCP is built on top of IP, TCP must therefore break your stream of data up into packets. So, some internal TCP code queues up the data you send, then when enough data is pending the queue, it sends a packet to the other machine.

This can be a problem for multiplayer games if you are sending very small packets. What can happen here is that TCP may decide that its not going to send data until you have buffered up enough data to make a reasonably sized packet (say more than 100 bytes or so). This is a problem because you want your client player input to get to the server as quickly as possible, if it is delayed or “clumped up” like TCP can do with small packets, the client’s user experience of the multiplayer game will be very poor. Game network updates will arrive late and infrequently, instead of on-time and frequently like we want.

TCP has an option you can set that fixes this behavior called TCP_NODELAY. This option instructs TCP not to wait around until enough data is queued up, but to send whatever data you write to it immediately. This is typically referred to as disabling Nagle’s algorithm.

Unfortunately, even if you set this option TCP still has serious problems for multiplayer games.

It all stems from how TCP handles lost and out of order packets, to present you with the “illusion” of a reliable, ordered stream of data.

How TCP implements reliability

Fundamentally TCP breaks down a stream of data into packets, sends these packets over unreliable IP, then takes the packets received on the other side and reconstructs the stream.

But what happens when a packet is lost? What happens when packets arrive out of order or are duplicated?

Without going too much into the details of how TCP works because its super-complicated (please refer to TCP/IP Illustrated) in essence TCP sends out a packet, waits a while until it detects that packet was lost because it didn’t receive an ack (acknowledgement) for it, then resends the lost packet to the other machine. Duplicate packets are discarded on the receiver side, and out of order packets are resequenced so everything is reliable and in order.

The problem is that if we were to attempt to synchronize this using TCP, whenever a packet is dropped it has to stop and wait for that data to be resent. Yes, even if more recent data arrives, that new data gets put in a queue, and you cannot access it until that lost packet has been retransmitted. How long does it take to resend the packet? Well, it is going to take at least round trip latency for TCP to work out that data needs to be resent, but commonly it takes 2*RTT, and another one way trip from the sender to the receiver for the resent packet to get there. So if you have a 125ms ping, you will be waiting roughly 1/5th of a second for the packet data to be resent at best, and in worst case conditions you could be waiting up to half a second or more (consider what happens if the attempt to resend the packet fails to get through?). What happens if TCP decides the packet loss indicates network congestion and it backs off? Yes it does this. Fun times!

Why you should never use TCP to network time critical data

The problem with using TCP for realtime games like FPS is that unlike web browsers, or email or most other applications, these multiplayer games have a real time requirement on packet delivery. For many parts of your game, for example player input and character positions, it really doesn’t matter what happened a second ago, you only care about the most recent data. TCP was simply not designed with this in mind.

Consider a very simple example of a multiplayer game, some sort of action game like a shooter. You want to network this in a very simple way. Every frame you send the input from the client to the server (eg. keypresses, mouse input controller input), and each frame the server processes the input from each player, updates the simulation, then sends the current position of game objects back to the client for rendering.

So in our simple multiplayer game, whenever a packet is lost, everything has to stop and wait for that packet to be resent. On the client game objects stop receiving updates so they appear to be standing still, and on the server input stops getting through from the client, so the players cannot move or shoot. When the resent packet finally arrives, you receive this stale, out of date information that you don’t even care about! Plus, there are packets backed up in queue waiting for the resend which arrive at same time, so you have to process all of these packets in one frame. Everything is clumped up!

Unfortunately, there is nothing you can do to fix this behavior with TCP, nor would you want to, it is just the fundamental nature of it! This is just what it takes to make the unreliable, packet-based internet look like a reliable-ordered stream.

Thing is we don’t want a reliable ordered stream.

We want our data to get as quickly as possible from client to server without having to wait for lost data to be resent.

This is why you should never use TCP for networking time-critical data!

Wait? Why can’t I use both UDP and TCP?

For realtime game data like player input and state, only the most recent data is relevant, but for other types of data, say perhaps a sequence of commands sent from one machine to another, reliability and ordering can be very important.

The temptation then is to use UDP for player input and state, and TCP for the reliable ordered data. If you’re sharp you’ve probably even worked out that you may have multiple “streams” of reliable ordered commands, maybe one about level loading, and another about AI. Perhaps you think to yourself, “Well, I’d really not want AI commands to stall out if a packet is lost containing a level loading command – they are completely unrelated!”. You are right, so you may be tempted to create one TCP socket for each stream of commands.

On the surface, this seems like a great idea. The problem is that since TCP and UDP are both built on top of IP, the underlying packets sent by each protocol will affect each other. Exactly how they affect each other is quite complicated and relates to how TCP performs reliability and flow control, but fundamentally you should remember that TCP tends to induce packet loss in UDP packets. For more information, read this paper on the subject.

Also, it’s pretty complicated to mix UDP and TCP. If you mix UDP and TCP you lose a certain amount of control. Maybe you can implement reliability in a more efficient way that TCP does, better suited to your needs? Even if you need reliable-ordered data, it’s possible, provided that data is small relative to the available bandwidth to get that data across faster and more reliably that it would if you sent it over TCP. Plus, if you have to do NAT to enable home internet connections to talk to each other, having to do this NAT once for UDP and once for TCP (not even sure if this is possible…) is kind of painful.


My recommendation then is not only that you use UDP, but that you only use UDP for your game protocol. Don’t mix TCP and UDP, instead learn how to implement the specific pieces of TCP that you wish to use inside your own custom UDP based protocol.

Of course, it is no problem to use HTTP to talk to some RESTful services while your game is running — that’s not what I’m saying. A couple TCP connections running while your game is running isn’t going to bring everything down. The point is, don’t split your game protocol across UDP and TCP. Keep your game protocol running over UDP so you are fully in control of the data you send and receive and how reliability, ordering and congestion avoidance are implemented.

The rest of the articles in this series show you how to do this, from creating your own virtual connection based protocol on top of UDP, to creating your own reliability, flow control and congestion avoidance over UDP.

Next: Sending And Receiving Packets

If you enjoyed this article please consider making a small donation. Donations encourage me to write more articles!

223 comments on “UDP vs. TCP
  1. Hey Glenn,
    great article – I learned the same via my academic studies, then developed a realtime fightergame as html5 with nodejs socket.io (TCP) for mobile devices and wondered why sometimes the characters werent moving, and if they were, they did all the movements, which i ordered before, in less than a second.
    now it makes all sense – i will either try to disable the Nagle’s algorithm or completely switch to UDP (but i guess not all browsers have implemented it yet).
    thanks, tho!

  2. Pingback: Good article on UDP vs TCP | Robi Sen's Blog

  3. Pingback: Optimizing Multiplayer 3D Game Synchronization Over the Web | How to play cool games

  4. Pingback: Lidgren.Network – explaining NetDeliveryMethod and sequence channels | GameDev<T>

  5. I’m publishing some video streams by RTMP protocal actually its application layer protocal . how do i find whether i’m using udp or tcp transport protocal to tranfer the packets..is there any shortcuts..?

  6. Pingback: Apa itu TCP dan UDP ? | Handitowira | Mari belajar dan saling berbagi

  7. What do you guys think about RUDP? The hybrid between UDP and TCP that checks for packet confirmations but doesn’t put as much pressure on the processor. It’s the recommended protocol for Photon Unity Networking and I’m using it for my lock-step RTS game.

  8. Thanks for the article. I have a question. For example there is an RPG game and the user pressed a button to use his skill. What if the pressed button information could not arrive the server?

    • For a situation like this if you were to use UDP you would implement a reliability system on top of it that would resend that message from the client to the server until it knew for sure the server got the message. This reliability system could be implemented slightly differently to TCP, so that on average the request to press a button would get to the server faster than it would using TCP, on average, if packet loss and latency exist between the client and the server — cheers

      • In this situation only client’s packets must be arrived reliably to the server. Server’s packets are not that important because if it fails, the next packets will correct the client’s data. So, client can send with TCP and server can send with UDP. But it’s only an idea :D, I don’t know much about network programming,

  9. this “article” is full of so much bullshit …. its actually fascinating.
    I’ll just name a few :

    1) TCP packets never “wait” on each other – not even segments do that
    2) applications which actually SYNCHRONIZE CONNECTIONS are … faulty and should avoided altogether
    3) TCP overhead is tiny – about 50 bytes / packet, if not less
    4) actually stating that games like mortal kombat cant be “done well” is just … hilarious, delusional and disconnected from reality
    5) in reality, TCP-packets usually are delivered EARLIER than datagrams – because they’re prioritized in every professional router / backbone, all of the ISPs all over the world do that; its an unwritten rule and can be seen in reality everday, even with the simplest of tools – this results in a better latency AND ping, although the difference is very small (usually .1 to 2 ms).

    my conclusion over the last years : either the amount if developer-incompetence is way bigger than i thought or some weird conglomerate pays wannabe-developers for spreading made-up bullshit and making the gaming-life harder for everyone. Im guessing the former option is what it is right now.

    I started routing my private WAN-connection over a VPN (over TCP) and ALL of my gaming has improved significantly since my VPN exit-point is a dedicated rootserver – these tend to be less restricted, even for unreliable wannabe-protocols like UDP.

    Protip for all the young want-to-be programmers : if you want to sound smart, use words like “UDP lite” instead of UDP … that’ll give you at least the benefit of doubt since that protocol has the least possible overhead 😉

    • Ahahah *wipes tear from eye*

      So when TCP packet N is lost, the next packet N+1 that arrives (out of order) before N is resent is delivered to the user immediately?

      Amazing. Absolutely amazing. I had no idea TCP worked that way. Bravo sir.

      • uh, yes – thats exactly how it works. Maybe you need to read up on the TCP specs, if not networking in general; plenty of “tears” are to be had. I think i know where you made your mistake (just a guess) : you dont know the difference between segments and packets. But thats no problem, we all need to start learning somewhere.

        While you’re doing your research : also be sure to look up on the expression “out of order”. It doesnt mean what you think it means.
        I’ll even give you some more aid : the only problem with TCP is its somewhat high CPU-usage, its certainly higher than with quite a few other protcols; but thats no argument anyore – with 2 cores or more that usage will be microscopic, compared to many other things.

          • Choice quote:

            “TCP is optimized for accurate delivery rather than timely delivery, and therefore, TCP sometimes incurs relatively long delays (on the order of seconds) while waiting for out-of-order messages or retransmissions of lost messages. It is not particularly suitable for real-time applications such as Voice over IP. For such applications, protocols like the Real-time Transport Protocol (RTP) running over the User Datagram Protocol (UDP) are usually recommended instead.[2]”

  10. “I’ve been developing high-capacity MMO servers professionally since 2008 that manage thousands, and in some cases tens-of-thousands, of simultaneous connections.”

    and you still need to learn how to read (palmface).

    good thing for you guys that its anonymous, lol 😀

  11. I an engineer at a voice over IP company – a competitor to skype. In any real time system where the customer bandwidth may be an issue – and when you consider a very large user base – UDP is far preferable, because the total number of packets transferred is lower. If you read the Cisco manuals carefully, as I have, you will see that they are more concerned with packets per second than bytes per second. That is the key measurement to consider. TCP will induce a much higher number of packets to be sent. If you are have your server in a data center, bandwidth is relatively cheap so the server connection to the backbone is not usually an issue. But the client side connection varies all over the map in the USA; we have fiber optic users with 50Mbits/sec connections, down to DSL connections which can be perhaps 0.3Mbits.. it takes more work to write a UDP transport protocol, but in the end a fancy UDP network protocol that incorporates retransmission of lost packets will end up creeping closer towards TCP in terms of the total number of packets sent to accomplish a certain task.

    Many of the massive multiplayer games have very complicated internal state, and it is awfully hard to program an update protocol that can work with partial of the message missing. So i would say, use TCP unless you need the ultimate in performance, in which case UDP is indeed the only sensible choice, but be prepared to reinvent the wheel, and handle every single possible malfunction case.

    Some games like Mortal combat fighting games or even Mario Kart type of driving games can’t really be done well multiplayer. It could take 50 Msec to get to a server say located in Dallas from California, and humans can feel 50 msec…. MIDI was a music protocol and originally it used a very slow communication channel, and if you put enough MIDI devices together, and triggered a chord on a bunch of machines wired together you could hear the smearing of the sound.

    The speed of light doesn’t turn out to be that fast, and this is why playing in the living room against a live person at 60 frames per second (17 msec between frames) is very satisfying. You won’t see PS4 or XBOX one disappear anytime soon, because they can keep that kind of frame rate. I doubut a FPS over the internet can do much above 15 fps in terms of actually responding.

    i think the conversation should have started with the key requirement: i want a certain latency in my system, guaranteed. TCP can’t guarantee a steady pulse like UDP, but for most games, high frame rate is not such an issue. I thought one of the comments about how when a multiplayer games becomes massively popular it starts to slow down is more likely caused by the fact that not every company can just go upgrade their bandwidth and buy new servers on such short notice, particularly when many games are free, and it is always a strain to do all this free stuff BEFORE you see any cash. Google has spoiled people because they have a money machine behind the scenes powering all this free stuff.

  12. I’ve been developing high-capacity MMO servers professionally since 2008 that manage thousands, and in some cases tens-of-thousands, of simultaneous connections.

    The author of this article is insane. I feel horrible for the developers you’ve mislead, and the companies that had to pay them.

    You should *maybe* consider UDP for low-latency physics, such as broadcasting entity rotation or finite mouse movement, but even this is unnecessary anymore with the abundance of broadband. It isn’t 2002. Anyone still playing from an AOL dial-up isn’t a consumer of your product.

    Telling people to NEVER use TCP is wildly, stupidly wrong, and runs counter to what is CLEARLY working across the industry.

    You’re telling people to bash nails into wood with their fists, because they might hit their thumb with a hammer.

    • Please learn to read:

      “The choice you make depends entirely on what sort of game you want to network. So from this point on, and for the rest of this article series, I’m going to assume you want to network an action game. You know games like Halo, Battlefield 1942, Quake, Unreal, CounterStrke, Team Fortress and so on.”

  13. I found this trying understand why after i have fix all the issues with my Connection having my IP provider come over & they replace all the line & even the Trunk line outside my door from the main box from the company. Yes i was watching them do it all. Bought new Router they give brand new Modem we put together our efforts to config everything just right This fix the Horrible drop conn issues I had. But Gaming 2 Days was running awesome & than back same shit Lag Lag extreme. Anyways this helps me understand is not me is just them.Will keep grinding this Bullshit till i balance somehow.

  14. Pingback: A work of netweek. Or close enough. - Fa Nacht

  15. Pingback: Seikun Kambashi | Facebook Hackathon

  16. Hmmm i do not understand all of the negatvitiy i have seen in some of the comments i have seen, this article is just stating clear facts and has been very useful for me, i am currently doing my dissertation on game networking and this article has provided very useful information, and as a bonus was very entertaining to read.

    Thank you for this information, i look forward to reading your other articles.

  17. Pingback: Good multiplayer networking approaches | Ohmnivore

  18. Pingback: NOMone

  19. Hi Glenn,

    thx for your answer.
    The only thing I don’t know how to achieve is the countdown for each move.
    So every player should have 15 seconds to decide what he’ll do.
    This behaviour should be handled on the server to avoid cheating and of course sync problems.
    Should I go like “send a get request to the server every second”?
    Seems stressy for the server to me. ..
    Additionally how to handle those push actions from the server like deal cards, announce winner, start new round and so on?

    In general there seem a lot of push actions located on the server ( like deal cards, announce winner and so on).

    • I don’t have a lot of web programming experience so I don’t know the best way to do this. One request or push per-second seems reasonable although I guess you could predict the countdown timer on the client while leaving the server authoritative over saying the final, ok time is up push event. cheers

  20. Hi,

    very interesting article.
    I´m just starting to develop some turn based games (e.g. Poker) and I wonder which network architecture fits best. I think sockets are a bit too “big” at all for a turn based game aren´t they?
    Would REST with JSON be an option? Or some kind of short/long pulling?
    I don´t think peer-2-peer would be a good choice because of cheating etc.


    • For poker I would say that a RESTful server is probably best, using web tech. You should have no problem with that approach, and no need for anything more complicated (eg. reliability over UDP etc.)

  21. Pingback: d5levelfc | [Compare]: TCP (Transfer Control Protocol) vs UDP (User Datagram Protocol)

  22. Pingback: Multiplayer using NoobHub, phase 1 | ALLINGAMES.

  23. I’m reading this article among the ones in physics, because I stumbled upon this after searching for networked fluids simulations.
    I have (not in industry) knowledge with fluids simulations and network games.
    The article has a strong author opinion side. However, I wish to share some experience gathered over experiments regarding this article, and personal opinions.

    It is rude and somewhat immature to insult someone who attempts to share knowledge because you think is wrong, regardless if he is you can point it out in civilized manner.

    UDP for action games (FPS, Racing games, etc) and also RTS, even with prediction, is far preferable to TCP. Also, multicast is know days somewhat more manageable in the application level. I have made a mini MMOFPS, implemented on top of JGroups, tested in 160 physical machines, in distinct LANs, with real players and AI, so far. In which the world is partitioned in squares, each corresponding to a multicast group, and players can interact in real time, without major asynchronous issues. A implementation using TCP, soon achieved lag issues. I have to point out also that the amount of data to update/communicate per players in fast paced actions games, is far greater that in RPGs.

    TCP however, for RPGs is often used. The reasons for such are:
    1-Because of TCP reliability.
    2-RPGs can infer a lot of information reducing significantly communication.
    3-RPGs updates can be smaller in amount of information to send than for fast passed actions games.
    Nevertheless, lag torments most MMORPGs in good days. What often happens is that after new games boom of initial players (sometimes over the initial predicted results) MMOGs tend to hold and survive or the crash into the ground. When they hold, after gaining a lot and losing some players, they achieve a somewhat fixes mass of players that will vary with some degree of predictability. This, is often well though but in many aspects unpredictable. World of Warcraft in its beginnings was in risk from shutting down. One of the oldest MMORPGs, Tibia, has many issues on this field, but to be fair WoW as a far better and younger network infrastructure design.

  24. Why can’t I use both UDP and TCP?

    This conclusion is misleading. The paper is said that burst tcp flow will cause udp packet loss. Of course. burst udp flow will also cause udp packet loss. It have nothing todo with mixing udp and tcp. The paper means that when network is busy, packet will loss that’s all

    • I agree. When you saturate your link you will induce UDP loss due to TCP behavior.

      Hence if you have a bandwidth intensive application (game) and it is running close close to total bandwidth on the link, then you would be wise to use UDP for all data intensive flows.

      Of course you could, if you must, send a small amount of data over TCP, just don’t send the bandwidth intensive part or your game protocol will suffer


    • I might be wrong but I think the point is that TCP will have more probability to clog network communication. Thus, it might (depending on what it is being used for) that TCP will clog UDP packets than the other way around.
      An example is a network where a few servers are suffering issues consecutively and network traffic is continuously being lost UDP doesn’t care, while TCP will resend packets a few times unless some TTL on attempts.
      You can for example have a chat client using TCP or a login service and the game package distribution being done via UDP for example.

  25. Great Read, but what is your solution to reliable data with UDP?
    Say we’re playing Team Fortress 2, you offers me trade for my amazing fez. Now the game becomes more of an IRC, suddenly we don’t care as much for the most recent data and rather individual events. Temptation to use a mixture of UDP&TCP!

    • Bottom line is it’s only really problematic to mix UDP and TCP if you are saturating the link. These days especially on PC you can totally get away with it. I’d still say though that really, it’s probably just easier to use a nice reliability layer over UDP for your game protocol vs. mixing separate TCP and UDP sockets. Also, NATs.

  26. Pingback: Introductory Research ← Dissertation

  27. Awesome article!

    Starting to make an MMO and thank god I read your article.

    I was concerned with performance using TCP over UDP, but wondered about the true differences and you could not of said it more intuitively.


  28. Thank you for the details, How you compare disconnection issue in TCP while in UDP since now active connection is required, can UDP be more reliable in chat applications?

  29. please …can any body tell me the percent of videos transmitted reliably to those that are transmitted unreliably(best effort) with the reference (paper or site you get the information from)

  30. Hi Glenn, Great article. (Actually great site, I’ve stumbled across it just now and been here 2 hours reading.)

    I’m not a network programmer but I found the whole protocol thing very interesting and it brought up a question. Would or could lag look differently to the client depending on which one was used? I would imagine that (at least with significant packet loss) the game state would appear to “snap” to the correct state with UDP but with TCP it would “fast-forward” as the reordered packets are resolved. Am I correct about this? Could this be used to diagnose where the packet lose was occurring in a system running both UDP and TCP?

    • Not really fast forward because TCP buffers out of order packets late tend to clump up and get delivered on the same frame.

      So they both look like a snap but the period of no packets arriving is longer with TCP than with UDP under packet loss

  31. I’m making a prototype for an online Bomberman type game (max 4 players for room), is TCP suitable for this kind of game ?

    Thank you!

  32. But, major browsers still only using TCP. (I know real-time connection is being developed in browsers)
    It must be non-FPS game, a first popular HTML5 game.

    FPS games are too main stream though, and it’s now boring.

    Brace yourself! HTML5 games, 3D web apps, real-time ‘internet’ are coming! even on mobile devices. And Javascript is rising!

    UDP vs TCP? it is relative.
    The best way is to know when both pros and cons occur.
    decide and *DO*

  33. Very informative article that helps to better understand why Secondlife is moving from UDP to TCP (http) to resolve increasing inventory and rebaking issues. I hope they will stick to UDP for other things, even I’m sometimes in doubt UDP is best as I experience increasingly a drop in network quality for end-to-end communication over large distance (from Europe to the servers located in the US).

  34. Love the articles. This stuff is so interesting to read about.

    I’ve been doing a little bit of networking with Unity, but I want to take more control and do some experiments of my own without using a pre-existing network solution, so this is really good to learn from.


  35. Pingback: Internet multi-player games: Server technology | NOMone Blog

  36. hi Glenn Fiedler
    i write a pachisi game that work on network, you right about tcp. i used it because i did not have a lot of input and output. sorry! i right can not speak english!

  37. Hello,

    I have only TCP checksum offload and UDP checksum offload in my lan, so i disable TCP and leave UDP enabled?

  38. I found this information very useful and informative.

    How does UDP get through routers without manually forwarding ports? I understand that it’s not really the topic of discussion here, but I feel that someone here could provide a better explanation from a developer perspective, then most google searches provide.

    Thank you.

  39. TERA uses only TCP for network communications, which you can verify using WireShark — it’s not against the Terms of Service. Full disclosure: I helped publish TERA but didn’t code the game itself.

    And incidentally, as the Guild Wars programmer who wrote the client->server communication protocol, the server->server communication protocol, and the low-level network libraries (among other things), I want to caution folks about making assumptions about why we chose to use TCP for Guild Wars (or why Blizzard chose TCP for WoW). We didn’t pick TCP for Guild Wars because it was unequivocally the best choice; it was a trade-off.

    It probably warrants a blog article, but the reason Guild Wars uses TCP is that I felt the labor required to write a reliable transport layer on top of UDP would take too long: we’d run out of money fooling around writing transport code instead of writing a game.

    So I chose to write Guild Wars on top of TCP, and if it didn’t work out I could rewrite the client transport layer to use UDP. Fortunately TCP worked well enough for client->server code.

    But there is a big downside to TCP that I should mention beyond what Glenn talked about: you lose control of when connections terminate. So if a router fails-over in your datacenter, or if one of your bandwidth providers fails, you’ll likely drop all your TCP connections traveling over that route. With UDP it’s possible to just ignore the loss of connectivity, wait for routes to re-converge through a different router or bandwidth provider, and keep going.

    So in short, Glenn knows what he’s talking about!

  40. There’s something I’ve been wondering about for a while.

    A hybrid Action-MMORPG called TERA Online is nearing release, complete with the tagline “True Action Combat”. I’ve played in the public beta tests, and can personally attest to the legitimacy of the claim; the controls feel crisp and responsive. What’s even more interesting is that it is one of the first ‘action’ games to feature a persistent world, a la World of Warcraft, Guild Wars, and other more traditional MMORPGs. No lobbies, no rounds.

    I did some googling, and it sounds like TERA uses both TCP and UDP to pull it off. Maybe I’ll try capturing some packets with Wireshark later (after double-checking the ToS to make sure I can do so) and take a look, but in the meanwhile, any thoughts on this?

  41. Hi Glenn,

    I am new to online Game development and I am trying to develop an online card game using ASP.Net MVC. I was reading your article on UDP vs TCP and some of your comments. You have explained it well to get basics of how both works. I understand for some games UDP will be the best choice. I am thinking to use TCP for the card Game. Is it a right choice and do you have good pointer which can lead me to right path.

    Thanks and I appreciate your reply.

  42. Hello Glenn, thanks for this tutorial and I will keep reading alone. Recently I am planning writing a simple multiplayer game demo, in which people can run against each other. I feel there is not too much data to be sent and the data does not need to be “very realtime”. So TCP should be good for me, am I right?

  43. Hello Glenn, a few days ago I started to search what will be the best (UDP or TCP) for a simple 2d online fight game in which two player fighting, please tell me what you think?

    Thanks! :)

  44. I will develop an online fps game.The client side must be developed in flash (actionscript).and flash player only support tcp. Is there any way to use udp in flash? If not, what can I do

  45. Networking is completely new to me, but I do understand your explanations (Thank you for that). Can you quickly explain the different in information between an MMO and FPS. I’ve always thought MMO’s were fairly quick paced games as well, at least when it comes to PvPing.

    Would they change the way packets are sent if a player was in PvP mode?

    Any help would be great! Thanks again for this!

  46. Hi, I am currently using the Quake3 model for my networking.
    When the client joins the game, it has to download scripts, ini files and a map if it doesnt have it. TCP would be the best solution here wouldn’t it? but if the server is sending lots of TCP data to a client is joining, while sending UDP data to the rest of the joined in-game clients, what effect will this have?

    Thanks :)

  47. In regards to what was written about using both TCP and UDP and the potential effect TCP has over UDP, will TCP traffic in other applications affect UDP traffic on the game application?

    • yes. my comment applies primarily to consoles — my point is that if bandwidth is tight you are better off having complete control over the stream. this is likely not to be a significant issue for anybody developing on PC — except, that you should always remember that somebody may steal all your bandwidth (bit torrent, video streaming etc…) and you need to degrade gracefully. cheers

  48. I only googled to make sure I haven’t been lying to people about UDP and I find this mess. I’m sorry Glenn these people are bugging you on protocol. I mean think of how awesome VOIP would be if they used TCP eh? I bet it would have failed hard. In network hierarchy I know that TCP is better and more reliable, but when you need bursts of data that software can make heads or tails of faster than a NIC card….. the answer is obvious. I’m only a Cisco nerd, I haven’t played in much game design other than the UDK. So, I’m speaking only from a network perspective but I think and know you are dead on Mr. Glenn. Hopefully this ignorance doesn’t bother you 3 more years and thanks for the write up.

  49. Pingback: Drawing in Flash using iPhone « Specialmoves Labs

  50. Glenn Fiedler,

    What I understand from reading your article and the replies above TCP stalling the network (due to packet-loss/resending packets) is quite limited when there is sufficient bandwidth available. So in relation to what you said about mixing TCP and UDP in a networked FPS, and assuming the total bandwidth usage of the game is quite limited, what is your opinion on the idea of using UDP for non-critical packets (eg. mouse movement/player position), and TCP for the more important (and fewer) packets that may affect others players more drastically (eg. mouse clicks/shooting)?

    • I think that if the reliable events you wish to send are time critical, then you would be better off embedding them in a custom UDP protocol and resending them at high frequency until they are acked — this works best if these events are small, relative to the stream of unreliable data, and infrequent.


  51. Mmm. I think this all debate of TCP vs. UDP is a bit out of context most of the time. It seems pretty obvious that, as Glenn mentioned, in fast paced games that you wanna network, latency is key. And by latency, I do not mean the quality of the line (of course it matters anyway) but the “application” latency.

    I have the feeling that these discussion tends to forget the initial and authoritative requirement : The network engine itself.
    Again, most of fast paced games (FPS like Quake for instance) network engines rely on the “last state is the right state” paradigm. Theses are client/server based most of the time to eliminate cheating (with an authoritative server) which leads to other interesting discussions between peer to peer vs. client/server based topology.

    Let’s assume you network engine is “state based” in a client/Server topology : your server main goal is to convey this last game state to clients and clients to provide “change” events (most of the time player inputs) to the server.

    So this is the starting point. given that context, you have 2 technical solutions to exchange data: TCP which is reliable (no packet loss, in-order delivery, etc.) and UDP which is not.

    It seems pretty obvious that, in that case, as you do NOT require reliable communication, you would pick UDP over TCP. You can on the other hand use TCP if you want to, but it’s just overkill as you don’t need it.


    • It gets interesting when you require mostly unreliable-unordered state, but need to deliver some reliable-ordered data, especially when this reliable-ordered data is time critical.

      • Yeah, you’re right. However, according to what I’ve read (and experiment) so far, requiring a reliable & ordered data is uncommon.
        I haven’t come to see any examples where it is required. I’ve read some of the game network engines articles available on the internet, like Q3 networking, Source (Half-life and others) networking, Age of empire networking and they all rely on UDP as far as I understood.
        However, I would be curious about a network engine requiring a reliable in-order communication somewhere in the process.
        Do you have any pointers ?

  52. I hardly ever add comments to blog posts, but I must say Glenn, your article is very interesting. I’ve been meddling with the issue myself for the past some 15 years of programming, and although I am by no means close to an expert, I would say what you’ve wrote is not too far from the truth. I am not a game builder, but I tend to deal with conferencing stuff a lot, and they require some sense of realtime transmission. True, Skype uses TCP and it performs better than UDP based protocols like H.323, but when it comes to realtime video, I would say UDP is pretty much the choice. Amount of data you need to process even with fancy codec like x.264, still places tough burden on TCP and trust me, NO_DELAY means nothing when it comes to transmitting 1024×768 in realtime. I guess in most MMORPG, you’re actually transmitting control packets or at some chunk of metadata (please correct me if I am wrong) and I think such can be accomplished with TCP even in realtime. But data transmission with 3Mbps and 20-30 fps, I don’t know what I would do without UDP. Of course, this is based on my limited experience, so if anyone knows how to implement such using TCP, I am more than willing to learn. :)

    • Yes you are absolutely right, MMOs tend to send control data or commands in reliable/ordered form. The amount of data is quite small and TCP handles this just fine. Action games tend to act much more like streaming video, exchanging a delta compressed unreliable-unordered series of state snapshots of the game world, therefore UDP.

      Of course even for action games you need some reliability *but* we want explicit control over the bandwidth stream up/down, hence a custom reliability protocol inside UDP.


  53. Hello Glenn I want to say I enjoy your work on network programming, it’s clear and concise. Had a few questions I was hoping you wouldn’t mind looking over if you have the time? I’ll do one at a time rather than a large bulk 😛

    When it comes to TCP and UDP, am correct in thinking that TCP transmits data similar to a continuous stream as opposed to UDP which operates on the basis of discrete packets? Let me try and illustrate. At a higher level we speak of “packets”, “events” and “messages” which loosely fits with UDPs definition in that as a protocol it preserves the boundaries of data you send and receive. You send 128 bytes of data, you receive exactly 128 bytes of data (generally speaking…). When it comes to TCP however, it appears to me (please correct me if I am wrong though) and based on papers I have studied that through various TCP mechanisms data could be aggregated or combined prior to sending? So if I send three unique segments of data, two of them may be combined meaning I am responsible then as the receiver to identify and separate that data upon receipt.

    Doesn’t that then lead to additional overhead in identifying the boundary of the “packet” I have received?

    • yes. as TCP you cannot rely on any “packets” you send coming through as whole packets on the other side, it will aggregate packets however it likes — you need to include your own packet delimiters in your TCP stream if you want to treat it like packets and this will create some small overhead


  54. Hello Glenn and thanks for a great article.

    I was wondering if you have any thoughts about possibly using SCTP instead of TCP/UDP for online games in the future? It seems to have several inbuilt features which could be taken advantage of, such as partial reliability, the ability to send both ordered an unordered data and multistreaming.

    From what I’ve read in articles the default settings for retransmissions and congestion control does hurt latency but since it’s a new protocol there should be room to work with and I’ve seen modifications suggested and experiments run which seem to suggest that it could rival UDP for latency without giving up any of its advantages over TCP.

    Here’s some interesting articles on the subject:

    • Yeah my approach to this reliability in this series is based around the assumption that you have a steady stream of carrier packets, eg. 30pps up and down continuously. Clearly this is not the case for MMO like games. In these you want to minimize bandwidth and resends, but in my custom protocols I want to minimize latency (at the expense of more bandwidth/redundancy).

      If I had to operate in a large scale, or high bandwidth cost situation, I’d consider RUDP, SCTP, or just plain old TCP.


  55. If you are just starting, I’d recommend using a library like OpenTNL, ENet or RakNet to handle the packet transport and reliability for you. These libraries work on top of UDP giving you the benefits of a custom protocol over UDP, without you having to write it yourself. They also handle NAT punch-through, which is important for P2P games to make sure everybody can connect from behind routers/firewalls.

    Once you have this working, each player just has to continuously send packets containing the position of their paddle, and the position and velocity of the ball, as they see it. Then, on each players machine, cross-fade between your view of the ball and the remote view of the ball according to the horizontal position of the ball (on your machine). This is how latency is hidden while also seeing the ball connect with the paddle of each player when it is hit.

    You can find more details on this classic method of networking pong here:


    • Hi

      Thanks For the reply.I want to know about the 3G network.
      As i am trying to play a game (Application on Android ) on 3G network Suppose i am moving from one 3G cell to another,Will the Ip Address change?? If Yes the How Can i continue the same game while changing the base station as my IP address is already registered in The game server.
      Will Handsover help me to keep the same IP .
      Is there something like DHCP.???

  56. Hello Glenn ,Nice Article.. As i am very new to game programming , I want to know few things about it.
    Suppose there is pong game application in a mobile(anderoid or iphone) and the game should be played between two person( on their own mobile) on the internet.While playing both the player can see the movement of paddle and ball of the other player.
    Can U tell me which protocol(Http , TCP , UDP) should be used to design this kind of game and Why?.What will be the latency of the protocol.

  57. Wow, this stuff is GOLD. I’ve been looking for information like this on the internet for a while, but only found your site right now.

    Thanks very much for putting in all the effort to write all of these articles. So far I’ve pretty much only used TCP (working on old-school text-based MUDs), but right now I’m designing an action-based multiplayer game, and UDP appears to be a great candidate for the transport layer.

    I’ve noticed delta-compression is used by engines like Source to reduce the amount of data sent per server snapshot. What techniques would be good to deal with dropped snapshots?

  58. Is there a typical limit to the number of UDP messages receiveable by a typical game server?

    Glenn, nice information you’ve got.
    Now, I would like your advise on something I plan to test.
    I understand I could setup a virtual connection over UDP. Since UDP can behave both as a client and server (depending how you implement it); I am wondering if there is a limit to how many virtual connections I can implement on a typical server (windows server eg.)
    Even if I were to implement a single or a pool of asynchronous thread/s to process the UDP message queue for all virtually connected clients; would there be some limitations to how many connections the server can handle for a real-time game. (Assumming each packet is 64K or less.)
    Would such a concept be able to handle 512 concurrent users easily? I imagining the load of handling all 512 messages in the message queue per game tick (30 times per sec. including game logic to process) Let’s say the typical user’s device is having a download speed of 500kbps and upload of 256Kbps. The server’s MTU is 1492, upstream 1024kbps, downstream 13997kbps.


    • Question is too underspecified to answer. The real question is whether the CPU cost of processing your UDP protocol is worth it. In other words, as you scale up what is the cost per-user? Do you still make money? Obviously this depends on your business model and the power/cost ratio of your servers.

      I’ve heard that many MMO and higher player count games use TCP simply because network cards run the whole TCP stack in hardware hardware and free up CPU for the game, allowing more game instances per-server, or more users per-server.

      I’ve also heard that MMO guys setup proxies for users connecting via TCP so that therefore any packet-loss is only between the user and the proxy and is quickly handled with minimal resend time. Practically speaking packet loss between the proxy and the MMO server on the modern internet is zero, therefore this solves the major problems of using TCP vs. UDP for games, while still being able to scale up to MMO scale.

      And of course it makes compression much simpler, you just run gzip over the TCP stream…


  59. INTRO
    Guys , last semester I did literature review about “Selection of Transport Protocol for MMORPGs” . I went with the conclusion saying UDP + TCP is the best choice for MMORPG. For that assingment I found some particular characteristics from two different papers (1 about Shanzhou online “On the challenge …. Some thing” and 2 I forgot but the main author was same). In both the papers characteristics were same. On top of that I found one more paper about GTP. GTP was discussing the same characteristics as well in order to prove that there is a need of protocol that is mix of UDP and TCP.

    Now Questions
    (1) Am I completely wrong?
    (2) Does such GTP exist?
    (3) How about using P2P architecture for such MMORPG (Particularly for solving scalability issue)

    Any new idea guys ,, let me know
    I was wondering if GTP exists and really used.

  60. Pingback: Michael Goldborough, Game Programmer » A Study of Transport Layer Protocol Utilisation in Game Design

    • You could go either way, in fact that’s what I’d offer — both TCP and UDP to the user. UDP with a simple reliability system built on top, and TCP so that users behind firewalls and with aggressive UDP filtering can play. Given that you’d have a deterministic lockstep model, either way is fine — the UDP solution would end up being a 100% blocking waiting on a lost packet to be resent, just with UDP you could pre-emptively resend data redundantly to reduce the delay waiting for resend that you would get in TCP.


  61. As I followed the discussion between TCP and UDP I have to back Glenn’s point of view up. The usage really depends on the type of game you want to create.

    For action based games like FPS you will always use UDP. Raknet for example provides you with a protocol on top of UDP in order to provide reliability as well. Or you can even write your own protocol as Glenn suggests. UDP is also a good idea here because in FPS games for example join short matches around 15min maximum. Multiple matches form a game session. These sessions have a periodical behavior which results can be reproduced.

    MMOG gamers will play one continuous session with varying parameters, this impacts the traffic patterns. MMOGs motivate to go online regularly to interact with the virtual environment. Therefore the inter session times may show a periodic profile. Due to some new circumstances, the maximum tolerable delays are much higher than those known from action games for example. As Glenn pointed out, action games that have a realtime requirement use UDP because retransmitted packets will have too high delay and will be discarded anyway. But for MMOGs like WOW TCP will be fine because larger delays do not cause problems. TCP is connection oriented and offers a reliable connection which will help preventing error propagation during long sessions.

    A new way which may cause the gaming industry to change in the future is the protocol SCTP, which combines the good things of TCP and UDP. The protocol can easily be integrated into the existing network infrastructure as only server and client have to know the protocol. There are already some studies regarding the fitness of this protocol for multiplayer games.

    However, really great articles.


    • I was thinking the entire article: why does he not mention sctp?

      You could just take a userspace implementation and integrate it into the gameclient and server.

      Another thing which I believe will have a bigger impact: zero-copy network traffic. Any packet is copied 3-4 times in inside a machine before it leaves the machine. This means most network traffic these days is limited by memory bandwidth. What they have come up with is dma for the nic(or rdma: http://en.wikipedia.org/wiki/RDMA). Quite nifty.

      I have bought 2 cheap leftover Infiniband adapters and one cable to have a fast interface between my little server and my pc. And fast it is. No difference between a local ssd and one in the server. Latencies are cut immensely(we are talking nanoseconds instead of the usual 5 milliseconds or so) as well. Rdma is the next big thing, and I hope a fitting protocol will come along with it.

  62. RE: Geo

    Dedicated servers generally run at fixed tick rates, but each player is in their own time stream (fps), and thus advance each player forward independently in time with (potentially) variable dt.

    Clients typically run at whatever framerate they like, the exception being if you are synchronizing via deterministic lockstep therefore each player would need to have the same DT for everything to stay in sync, also, fixed dt (or at least, a multiple of a fixed base dt) is very useful for networking physics simulation


  63. Hi

    On UDP, can you guarantee that one recvfrom() matches exactly one sendto() ? I.e if a client sends 3 messages/frame at 30fps will they arrive in 3 recvfrom/frame? On TCP they will surely arrive concatenated by the TCP layer. The problem is code design for parsing the array received from the socket – if i expect a single UDP header(seq, ackbitfield) and more than one game headers( a server sends X*playerAction messages concatenated by me) ; or i must expect multiple udp headers.

    For an action game, is it ok to have always 2 sendto/frame ? i.e. one for position(unreliable) one for actions(reliable) or the number of sendto/frame shoul be minimal?

    Thanks a lot!

    • no. you cannot assume that packets are grouped together per-frame as they were sent. they may be in practice mostly like this on the other side, but consider that sometimes packets will be late, out of order, duplicated etc. you’ll see as well obviously packets could be jittered, or stretched out in time a bit, or compressed together a bit…

      so basically if you want to send multiple packets per-frame, and they are all required for that frame of data, it would be much simple to just make them single packet.

      this way, you’ll either get the data for that frame, or not — instead of having 3 or so packets to track loss on

      also, you’ll save on bandwidth — each UDP packet has ~28 bytes of header overhead. 3 packets = ~56 bytes wasted vs. 1 packet with same data


  64. Pingback: [NRPPJ - 1/12] UDP vs. TCP » BloGUI

  65. Hi,

    When developing a multiplayer action game over UDP,i guess we should have a max message size that is smaller than the MTU. If this is true, for an online game played on wireless devices a safe max message size would be 512?

    Thank you, for the articles and for your time! :)

      • Thanks.

        I am curious about something. You always refer to 30 packets/second so i am guessing that all multiplayer games must be tick based(i.e on tick X everything is the same on all machines).

        The problem is that on slower machines(phones) you cannot limit the fps to it;s lower value because it would be too small. A way around it is to have constant 30 calls/sec to update and up to 30 calls to render. So each render will interpolate between the next update and the last one. Another way is to use everything time-based and have variable fps between machines. I think this may be a problem implementing lock-step games… what do you think?

        Thank you again for sharing your knowledge!

  66. Mr. Fiedler, I had to chuckle a bit at the “Flash sucks” comment simply because it doesn’t support UDP :)

    To be fair, only one of the three ‘major’ cross-platform browser plugins for gaming does support UDP, and that is Unity. Silverlight and Flash only permit TCP. Of course Flash is absolutely the dominant of the three, having something stupid like 99% market penetration.

    There is some work on UDP protocols for Flash (eg., Adobe Stratus,) but as far as I can see they have no intention of exposing direct UDP socket control.. yet.

  67. I would say that point #4 sums it up.

    You can pretty much use either TCP over UDP if your client side prediction is good enough.

    But, you have to weigh it all up, I guess.

  68. Hi,

    I don’t understand why you discourage the mixing of TCP and UDP flows, even I agree when the bandwidth is limited (old PSTN modems), there can be issues.

    But nowadays most of users have sufficient bandwith. For example, I often watch a TV flow (using UDP) while browsing the web and/or downloading (TCP) at the same time. As long as my bandwith is sufficient, there is no problem of packet loss.

    Why would it behave differently in a networked game, which never consumes lot’s of bandwith compared to nowadays connexions ?

    • The “only” means “for your game protocol” use UDP only. Not that you should not use TCP *at all*.

      An example as to why I wrote this originally: on the 360 you had to show that your game is capable of running at 64kbit/sec up and down without disconnecting the user. If you mix TCP and UDP flows for parts of your game protocol then this is really hard to do because you aren’t in full control of the bandwidth usage of the TCP part.

      There is no problem having small parts using TCP (eg. games using http to matchmaker services, cloud services is quite common these days), but putting large portions of the game protocol itself over TCP means that you are no longer in control and it’s difficult to get close to the available bandwidth.

      Whether or not this is actually a problem depends on what you are targeting, but with UDP only you do degrade better when other stuff is saturating the link (eg. somebody else watching netflix over your router) while you play a game. This is actually a common situation — to be bandwidth constrained to something like 100kbps even though the link itself is theoretically capable of 256kbit sec.

      So I think this advice is still valid in 2014.

  69. then, why a AAA+ company like Blizzard decided to use TCP for a 10+ millions player’s game if TCP doesn’t fit realtime game requirements for their MMORPG ?
    I suppose they had enough knowledge to implement a protocol over UDP. If they decided to use TCP, they must had good reasons to do so. Isn’t it ?

          • No offense, but I dont think you have more knowledge of network programing than Blizzard.

            WoW is a real time game, you have a cooldown between abilities, ok, you can’t shoot 30 missiles per second, but that is not the definition of a real time game… Time reaction in PVP are the base of the game.

            I’m a multimedia engineer, my final grade project was about real time multiplayer systems, implementing a server system and a videogame demo like world of warcraft arena system (in TCP). I know what i’m talking about… I’ve seen the pros and cons of TCP and UDP.

            • Dude. This is a series of articles about how to network a real time action game. Eg. FPS

              If you want to make a 2004-era MMO using TCP just go ahead and STFU

          • “No offense, but I dont think you have more knowledge of network programing than Blizzard. ”

            No offense but you clearly have to do your homework before you say things like that. Glenn Fiedler is pretty well known in the games industry, and is the lead network programmer at a pretty prestigious game studio. So just because you have played a video game, and done a school project on video games doesn’t mean you know more than an industry veteran.

            Also you’re a “multimedia engineer” and doubting the pros of UDP in this situation? Are you going to tell us you would use TCP for live video streaming now?

          • “WoW is a real time game, you have a cooldown between abilities, ok, you can’t shoot 30 missiles per second, but that is not the definition of a real time game… Time reaction in PVP are the base of the game.”

            It’s one of the method to mask the latencies of TCP. But in FPS games, you can’t do that.

  70. They probably do have a good reason, but this article is not about how to write an MMO, it’s about how to network action games.

    From the introduction to this article:

    “The choice you make depends entirely on what sort of game you want to network. So from this point on, and for the rest of this article series, I’m going to assume you want to network an action game. You know games like Halo, Battlefield 1942, Quake, Unreal, CounterStrike, Team Fortress and so on.”

    So let’s stop arguing this point because it’s not even remotely related to the content of this article.

    In conclusion:

    1. If you want to deliver time sensitive data and do not want for it to be delayed waiting for resent packets, use UDP.

    2. If your data is not time sensitive and you don’t mind having gaps in your packet stream waiting for packets to be resent, then by all means go right ahead and use TCP.

    3. In flash you can use only TCP. Looks like flash sucks balls.

    4. Some MMOs use TCP (WoW), while others use UDP (Eve Online). It is conceivable but by no means conclusively proven that there may be scalability and/or simplicity benefits of using TCP over UDP for MMO games, given that Blizzard is not stupid, assuming of course that you have a good client-side prediction algorithm in place to smooth out the gaps when packets are resent.

    end of discussion :)

  71. That’s exactly why i commented this article: saying that UDP is mandatory for real time networked games is just TOTALLY WRONG. TCP can fit in most ACTION games (and not only turn based games) and will save a lot of effort for the developper for almost the same result.

    As a proof, I already mentionned several sample links of REAL-TIME GAMES with real-time actions and real-time interactions between players (omgpop.com, …). They use TCP (the only protocol allowed in flash) in their flash games: bomberman-like, mariokart-like, … which are far from being turn-based (solitaire)


    – NO TCP is not only for turn based games (is WoW a solitaire-like ?) ; it will fit most of real-time networked games too (again: except FPS).

    – YES TCP is internally a bit worst than UDP for this purpose, but in real life it is un-noticeable for the player and it saves a lot of efforts to the developpers.

    • I strongly disagree. If you have a realtime game, then by definition certain information needs to arrive as quickly as possible, for example object state updates in a FPS. If you use TCP as a transport when a packet is lost you must wait for old data to be resent before you can receive the most recent state which you are interested in.

      This problem can be avoided by using UDP instead of TCP and developing a simple reliability layer on top of UDP. I don’t believe that implementing a reliability layer is beyond the level of ability of a typical professional game programmer, but even if it was, there are plenty of libraries that will do it for you: Raknet, ENet, Torque etc.

      You point to flash games as some sort of proof of concept, and in the same sentence mention that it is the only protocol allowed in flash. Have you considered that perhaps the only reason these games use TCP is because they have no choice?

  72. Some points on this paper:

    – their tests assume a bandwith of a few hundreds kilobits/sec, which is clearly insufficiant and probably leads to packets drops. This is not the case in the real life where the backbone and game servers are typically well dimensionned to support the amount of traffic received and sent.

    – their tests have also been performed with a loss packet rate of 4% ! I never encountered such high packet loss rate in my life (except if I unplug my network card :o). Concretely, that means for their data sent 8 times per second (their testing base), something like one loss happend every 3 seconds ! This is even not coherent with their remark a few line before telling that some clients never encountered packet loss during the whole testing session.

    For sure, as the biggest advantage of UDP over TCP is when you encounter packet losses, their result can only give a big advantage on UDP compared to TCP. But such network condition are purely theorical and never met in real life. Packet losts are the exception, not the common behavior. And as long as there is no packet loss, TCP behaves globaly as well as UDP.

    > ” there must be some reason to use TCP in MMOs, and that paper seems to directly counter that argument ”

    Blizzard did it for WoW (I mean using TCP). And we cannot reasonably tell that Blizzard is a company which is technically incompetent. So they had inevitably good reasons for using TCP.

    Last but not least, about technical things which are not present in the paper:

    – UDP is not as well routable as TCP on the net, especially behind NATs, Proxies, Firewalls, … That’s a fact and one more reason that makes some ‘real time’ application embed kind of ‘degraded mode’ to be used (and working) over TCP when UDP is not routable. Skype protocols are a good example : Skype makes use of TCP in environments that don’t accept UDP connexions, and it works almost ‘seamlessy’.

    – UDP requires a LOT more technical knowledge to be used in a networked game than TCP. This is much more efforts to implement a complete unbugged network engine compared to TCP. The time you spent on this part of the project is time lost for other parts of it.

    The real questions are:

    – is UDP mandatory for my ‘real time networking needs’ ? I would say that in 90% of gmae projects (I mean almost all games except FPS like ones), this is not absolutely necessary. TCP will behave (very) well too.

    – If UDP is not mandatory, will UDP be so much better than TCP that it could be interresting to spend much more time on the network engine instead of other parts of my project ? The answer also depends on the type of ‘real time networking’ you need, but generally no.

    My 2 cents.

  73. Anon, I think you should rather make your own tests than only relying your judgement on articles readen among the web.

    You may want to check too already existing online action flash games like those of omgpop.com, or even WoW to see that TCP usability in arcade multiplayer games can be a choice too.

    Everything is not fully black or white.
    I think that’s also true for the limit between usage of TCP and/or UDP in games.

    My 2 cents.

    • I think it’s pretty fair point to look at that article and conclude that TCP is probably not a good choice. If you would like to refute the points raised in the article and the research supporting it – go right ahead, but I hardly think it’s fair to tell somebody to ignore research and “make your own tests” when the research is in fact doing exactly that… testing how TCP performs in a real world situation.

      • Not only the papers deals only with one specific type of game (MMORPG) whereas Anon concludes it’s the same for all multiplayer games, which is wrong.

        But also the paper only deals with the advantages of UDP over TCP, but this paper’s goal has neven been to give an objective comparison, as there are no mention about udp drawbacks.

        Finally, I already gave a few arguments in my previous post (9 oct), and I’m still waiting counter arguments 😉

  74. Some additional informations about Darrin West’s remarks:

    – there are several ‘action’ games made in Flash (so using only TCP) which do behave very well.
    Some good examples: the website omgpop.com which includes a bomberman-like and a mario-kart like game. I even saw once a flash FPS (2D graphics, top view, don’t remember his name) which was behaving ok too.

    – Another example: Skype, which falls back to a TCP based protocol when UDP is not suitable (behind a proxy, …). It is still acceptable even if voice streaming typically has ‘real time’ needs too.

    – Last but not least: in a client/server ‘star’ topology, TCP leads to an higher connexion success rate. For example, people behind a proxy have almost no chance to use UDP while TCP may be suitable (like corporate or university networks).

    To be clear: I’m of course not saying that UDP is crap and TCP is heaven: typically UDP remains the best choice for FPS games. On the other hand I think it’s quite important to not let beginners think UDP is the ONLY way to do things.
    A bomberman game, a car racing game, a FPS and a pong game don’t have the same requirements, and in some cases (when UDP is not an option like in flash, or when you don’t want to loose too much time in protocol design) TCP may be a really good compromise.

    Darrin West said that better than me. His two remarks (“It is “hard” to get all this stuff working right” and “Be very sure your latency requirements do demand all that investment”) are really important and must be carefully considered before making a choice.

    Best regards,


    PS: sorry for my bad english 😉

  75. I can think of a good reason to go with TCP over UDP, so perhaps your “never” could use a qualifier. Flash currently only supports TCP. So if you want to make a multiplayer game for Flash clients, you’ll need to do it over TCP (or HTTP which is of course even worse, speed wise).

    • that’s a fair point – but i’d consider it a great reason *not* to use Flash instead :)

      when do you think they’ll fix that? it’s madness – 2009, and you can only use TCP in your flash application? *boggle* :)

  76. but what about the problem with getting past firewalls? when a TCP server accept()s, can you have any control over the port number it chooses? if not, how can the client properly set up his router to allow port forwarding to his gaming PC? I know UDP can and must define the port for send and recv; so there’s no arbitrary port number assignments.

    If anyone can show me how to bind the TCP sockets on both ends of the connection (client and server), I’d be happy to try TCP.


  77. Some more thoughts on TCP vs. UDP:

    * In network protocol design, be careful to optimize for the “normal” case, not the error cases. Think about how often a packet is actually dropped, and whether it is your own fault. Do you want to buy into a lot of protocol design (no matter how much fun that might be) to support the few users that are stuck on a low bandwidth connection?

    * Surprisingly UDP tends to drop packets because the application uses more bandwidth than the user’s ISP connection can handle or if the app doesn’t drain the local buffers more often. Backbone WAN connections almost never need to throw something away. UDP packets are pretty much never lost on a LAN.

    * From your app, you can tune the TCP stack to not have such large delays in transmitting a send request removing a *log* of latency. If you don’t overload the connection resending isn’t as big an issue as you might think.

    * Look into TCP_NODELAY (http://en.wikipedia.org/wiki/Nagle%27s_algorithm)

    * It is “hard” to get all this stuff working right if you build a custom protocol over UDP: bundling (reducing header overhead), rate limiting, congestion control, packet splitting, prioritization, connection error detection, …

    * Be very sure your latency requirements do demand all that investment. It is a slippery slope. Start with latency hiding techniques (predictive contracts, optimistic client movement and fixup when the server responds…), they may wind up good enough. You may even be able to change your game design to not require so much finesse. Bioware survives very well just using dice.

    * Data centers prioritize TCP and UDP differently. WAN switches have a lot of different modes. You may find your low priority stuff accidentally choking the stuff you wanted at high priority. These WAN level priorities are different in Asia than in the US.

  78. UDP for FPS games – no question.

    A collection of possibly interesting papers and reports on various aspects of FPS games and their network traffic patterns can be found at http://caia.swin.edu.au/genius/ (also some papers on server discovery issues, and player tolerance of latency)

  79. agreed. the point i am making in this article is that you should not use TCP and UDP together for the *bulk* of your protocol data, for example lets say 50% of your data is reliable, 50% is unreliable — and together these sum up to something close to the available bandwidth …

    in this case, you can get a bit stung by your TCP vs. UDP, for example assuming TCP needs to do resends, so you can go over the bandwidth required, and you start inducing packet loss in UDP, then your TCP dials back of course — but you’ve now lost a packet that needs to be resent, so your reliable events are delayed

    the point is that this condition is entirely avoidable by doing everything inside UDP, because you can ensure that you have a flat bandwidth profile, say 256byte packets 30 times a second, that always stays under the available bandwidth – no matter what is going on, so worst case within your internal reliable protocol, some events are throttled and arrive a (tiny) bit late, but your bandwidth doesn’t increase under poor conditions (no resends)

    now, of course if you just had the occasional event sent via TCP, and the bulk of your time sensitive data in UDP, no problems at all


  80. Speaking as a network researcher and not a game developer, the conclusion to never use TCP and UDP together seems a bit strong. TCP will only have packet loss if it is sending too much data; in some ways just like the UDP data you are sending. The difference is you have no direct control on the rate TCP sends at, this is hidden to you.

    If you just need to send some reliable data and do not want to worry about retransmission and implementing a reliable protocol, and you know the rate will be low, then there wont be any problems with using both TCP and UDP.

    The relationship isn’t all that complex between the two really: TCP merely increases its send rate (if there is data to send) until it gets packet loss, in which case it dials back its rate, then starts increasing the rate again (this time more slowly). When its increase in rate causes packet loss, it is quite likely to hit any other streams of data as well, including your UDP packets.

  81. Reasonably few, I’d imagine. Very different beasts, with different latency tolerances in gameplay. I was just addressing the MMORPG question specifically :).

  82. “How many fast-action, no-mouse-click-based MMORPGs use TCP?”

    Quite a lot of them. Including World of Warcraft.

    Game-state in MMOs is a pretty complex affair, meaning you often just send incremental updates of specific bits of data for your game objects, and not sync out the whole state. Interdependencies of data can also grow pretty strong. For most data, reliable ordering is a necessity.

    Some MMOs actually do use both TCP and UDP, with most traffic going over TCP and things like position updates going over UDP. Dark Age of Camelot did this, for example.

    Quick list found via some Googling:

    UDP: Everquest, City of Heroes, Star Wars Galaxies, Ultima Online, Asheron’s Call, Final Fantasy XI

    TCP: World of Warcraft, Lineage I/II, Guild Wars, Ragnarok Online, Anarchy Online, Mabinogi, Age of Conan

    TCP + UDP: Dark Age of Camelot (and I’m going to assume Warhammer Online as well, due to being the same engine and company.).

  83. “That’s why all major MMORPG uses TCP…”

    How many fast-action, no-mouse-click-based MMORPGs use TCP?

    For a point n’ click MMORPG it’s great to use TCP, plus, you can always see people complaining about lag on the global chat 😉

      • Actually. A lot of popular games use TCP. And you know what happens? The players hate the lag. Then comes mods. Mods save the day for TCP games. Guess what. Leatrix Latency wasn’t made for WoW for no reason.

        • As a Guild Wars and Guild Wars 2 (WvW) player, I would very eagerly anticipate any new PvP-focused MMO that hired plenty of programming talent with extensive experience producing FPS games (concerning the rendering/input engines as well as the network protocols). ANet spent five years making their game, and the underlying gameplay mechanics are great but the practical experience is embarrassingly riddled with game-breaking horrors, especially regarding both network responsiveness /and connection reliability/.

          Most of the disappointing results are clearly either design decisions (like fundamentally flawed world-colliding camera mechanics, camera easing, and other input delaying “features”) or design/architectural consequences (like choosing TCP over UDP, and using an out-of-process browser engine for the trading post–which makes a compact interface unreadable because of post-render font resizing). TCP-based protocols are just an easily identified contributing factor to their blatantly manifest shortcomings, due to its (academically and practically) well-understood behavior, and its performance costs which are so descriptively equivalent to the results we users experience.

          Sure you can note that MMOs are more complex and deal with way more data and potentially more data-ordering sensitivity, but that’s exaggerating the difference, which is in scale, not kind–and to a lesser degree as FPS games also start adding skills, rpg elements, and larger player counts. Any MMO’s /can/ accomplish a level of responsiveness+correctness with today’s 1-100MBit connections and 20-40ms minimum network latencies that Q3A managed 14 years ago on dial-up with minimum latencies of 90ms.

          And until they actually do, as far as I’m concerned they’re doing it wrong. There’s just no way it’s “too hard.” At best they’re not willing to make it a constraint on the gameplay mechanics, but more likely it’s just not a priority at all. Given that ANet decided to house ALL their North American servers in one datacenter in the furthest possible corner of the continent, in their case I’ll wager it’s the latter.

          Anyway, my roundabout point: sure MMO’s use TCP–it sure as heck don’t make it the right choice.

  84. “That’s why all major MMORPG uses TCP…”
    My understanding of why MMORPG etc use TCP over UDP (this is from reading blogs of network programmers from these types of games) is that data is more important than anything else where as in a fast paced game (FPS for example) the most recent data snapshot is most important.

  85. yes absolutely i should add a small note about security and NAT, the idea would be to use a session id and a packet hash based on some secure crypto

    this also helps with NAT because some routers like to dynamically reassign port on the fly for some reason, i’ve heard


  86. @Glenn:

    Well laid out article, as most of yours seem to be.

    I suggest you remind people that using a client’s IP address is not a reliable token to base a session on.

    I mention this due to I’ve seen it overlooked many times before, resulting in a myriad of possible exploits due to the relative ease of session hijacking.


    Because it’s totally implausible that those developers most likely don’t have the resources/time/know-how to develop a custom transport protocol; right? Lets not disregard the fact a good percentage of programmers are just afraid of the unknown, for whatever reason and would prefer to just stick to standards like sheep due to their own ineptitude.

    • I wouldn’t say its due to their own ineptitude, but rather due to time constraints. I often want to try something new, but it takes time to implement something you’re unfamiliar with, and you’ll likely bang your head against a few walls doing so. Worst case, you could introduce a bug somewhere that could take hours to track down.

  87. Anon:

    UDP is standard technique for action games. All FPS and action games use UDP. Its such standard practice that UDP vs. TCP is always the first question I ask in a network programmer interview :)

    Not convinced? Ok fine, lets look at the games that use UDP then: tribes, quake, unreal, starwars battlefront, battlefield 1942, halo, gears of war, team fortress 2, COD4… yes, pretty much all of the top-tier action game use UDP.

    Now, regarding MMOs i have heard some of them use TCP fine and get away with it. But remember that MMOs don’t have the same characteristics as an online action game. I’ve not worked on an MMO, but I’d guess in this case they would have lower frequency updates based around a reliable-ordered stream of commands, whereas action games are based around high frequency updates of most recent state.

    Given that this article series is about networking action games and not about networking MMOs, I think we can all move forward in the knowledge that UDP is the right way to go, in context


    • if you have something such as a attack command (user fires a shot, swings their sword, etc), do you implement some kind of lossless UDP? Or, is their perhaps another model?

      • these things map well to reliable events. basically keep sending the event at some desired rate (more frequently = more time critcial info) until you receive an ACK for one of the packets which the event was sent in.

        read the article or reliability and flow control later on in the series for more info about how to setup an ACK system over UDP


    • Ha, you have a lot more patience than I have, Anyway thanks for the awesome article, im looking to do my honours game project using networking between mobile devices and this could come in really helpful!

    • I think you pretty much described it best already, what the difference is, and coming from the MMO/RTS perspective, udp is quite understandable as fast alternative for FPS, but you have to consider how many connections you will handle – in the end, a good async solution will spend less time in network and handle lags better in tcp, than in udp, and e.g. for deterministic simulations like RTS you would have to reimplement tcp in udp anyway if it gets too big and complicated.

      Therefore, i think the only failure the article has, is your strongly expressed opinion, down to the fact, you see that as a question asked at interviews. You will discover anyway, that opinions, while certainly being valid, sometimes lead to generalization.

      This upsets the readers, who do not share the opinion. And quite frankly, neither do I completely, but mostly.
      SC2 e.g. uses tcp pretty well for action based stuff too. it depends entirely on the needs of the network layer for a specific simulation, and a good packet system will be usable in both.

      Still, big thanks for these articles, I consider them very useful, especially since you cover the basics.

      • Arguably, for an RTS as well, I can make a strong case that UDP is better. Complexity of your own reliability over UDP is solvable given a talented engineer and some time. The end result is all that counts: which is, how well you are able to handle packet loss over latency before your game becomes unplayable. With RTS games this means you can send via redundancy and reduce latency do to resends added by TCP.


  88. Conclusion:

    That’s why all major MMORPG uses TCP…

    Did you ever develop a real time online game? If so, you’ll see that the TCP is good enough to make one and UDP will not add better performances compared to the time you need to rewrite a good protocol on UDP.

      • Do you even understand what he is saying? This guide is pretty fucking stupid, as are you for following it word for word. TCP is great, and unless you’re writing an FPS, it performs perfectly fine. My question is how low your IQ would have to be to listen to this guy when he says “Even with Nagle’s algorithm disabled, TCP still does bad fuck TCP” (paraphrased)? How does it do bad?

        • Turning Nagle’s algorithm off with TCP_NODELAY does not stop TCP waiting for lost packets to be resent – it just turns off the code that coalesces small packets into one large packet on the send side.


          When a packet is lost TCP must wait for that packet to be resent. While waiting any other (out of order) packets received are held hostage in a queue so that the data transmission looks like a reliable ordered stream. If this behavior is undesirable, then UDP is a good choice.

        • it may WORK but it is a waste

          you’re adding unnecessary crap to your packets with TCP just because you’re too lazy to implement exception handling and error checking

          UDP is basically do what you want, when you use TCP, all you are doing is adding an extra layer of software, it is therefore obvious really that since this is a game and not a fileserver, you should be using UDP.

          How any of you actually think there is a debate in this matter and this article is doing anything other than stating the flat out obvious, is pretty silly…….. HOWEVER… TCP may be treated as trusted by a firewall when UDP is not (filesharing)… this is why, and only why you might choose TCP in environments where latency is less of an issue (not FPS games).

          I note runescape uses TCP and our warehouse staff are always playing the bloody game. That’s a perfect example, no doubt the creators want it to work anywhere ANYONE can get the internet.

          • Hey Richard,

            Realized I’m coming in late to this conversation but I wanted to mention that what may seem obvious to you may not be to others. I found this article useful simply as a practical discussion of why one may or may not want to use TCP/UDP because most of my exposure to UDP AFIAK has been very academic.

            It’s useful that you bring up some other practical considerations that companies have to consider when they develop a game – like runescape being accessible behind certain firewalls.

        • I second what this article says, after trying to get a functional prototype and getting stucked with lag issues using tcp. Use udp.

    • You’re conflating the latency increase present in TCP with the time required to re-engineer certain custom TCP functionality on top of UDP. While every good coder should consider whether the time they’re spending on a feature is worth it, the two are not fundamentally inter-related. If TCP is making an FPS run choppy, and bare-bones UDP isn’t sufficiently stable, then it’s not just desirable but necessary to implement your own version.

    • Anonymous – I will ask you – what MMO do you work on that is using TCP?

      I can tell you I know of an example of the biggest MMO failure in history: Anarchy Online. Anarchy Online struggled for 2 years to fix lag issues, that were so extreme, players could not even play the game. The issues came down to two parts:
      1. AO used exclusively TCP
      2. AO had a client side memory leak

      The memory leak was fixed over a series of patches the first 6 months, but the latency become serverside. While the client started behaving better, the dev team on AO just couldn’t resolve it. It took members of the community who were part of other MMO’s, who liked AO’s concept, to look in and tell them what’s going on…

      The use of TCP – the turnaround on packets was too long. Pretty much everything this poster said in blog here was spot on the problem with Anarchy Online from launch to launch + 1 year. My understanding (and I could be wrong) is that the yresolved the issue by gutting TCP and goign strictly to a UDP model.

Leave a Reply

Your email address will not be published. Required fields are marked *