资源

课程

第十八节：网络游戏的架构基础

Online Gaming Architecture: Fundamentals

网络游戏架构

PLAY ANYWHERE WITH ANYONE

与任何人随时随地玩游戏

Game developers have never stopped exploring multiplayer online gaming

游戏开发商从未停止探索多人在线游戏

Challenges in Multiplayer Online Gaming

多人联机游戏的挑战

Consistency

一致性

Network Synchronization

网络同步

Reliability

可靠性

Network Latency

网络延迟
Drop and Reconnect

断线重连

Security

安全性

Cheats

作弊
Accounts Hacked

账户被黑客攻击

Diversities

多样性

Cross-Play

交叉游戏
Rapid iteration

快速迭代
Multiple Game Systems

多游戏系统

Complexities

复杂性

High Concurrency

高并发性
High Availability

高可用性
High Performance

高性能

Network Protocols

网络协议

The Founding Fathers of the Internet

互联网的创始人（一开始用于军事用途）

Designed the TCP/lP protocols andthe internet architecture.

设计了 TCP/IP 协议和互联网架构。

In 1977 Cerf and Kahn will link three networks (packet radio, satellite, and the ARPANET) and prove the efficacy of their TCP/lP protocol in a dramatic round-the-world transmission from a moving vehicle, the SRl Packet RadioResearch van.

1977 年，Cerf 和 Kahn 将连接三个网络（分组无线电、卫星和ARPANET），并证明他们的 TCP/lP 协议在移动车辆 SRl 分组无线电研究车的戏剧性环球传输中的有效性。

How to communication between two PCs

A and B must agree on the meaning of the bits being sent and received at many different levels, including

A 和 B 必须就在许多不同级别发送和接收的比特的含义达成一致，包括
- How many volts represents a 0 bit, and for a 1 bit?
  
  多少伏特代表 0 位，多少伏特代表 1 位？
- How does receiver know which is the last bit?
  
  接收器如何知道哪个是最后一位？
- How many bits long is a number?
  
  一个数字有多少位长？

The Problem of Communication

沟通问题

Re-implement every application for every new underlying transmission medium?

为每个新的底层传输介质重新实施每个应用程序？
Change every application on any change to an underlying transmission medium?

在更改基础传输介质时更改每个应用程序？
No! But how does the Internet design avoid this?

不！但是，互联网设计是如何避免这种情况的呢？

Solution: Layering

解决方法：分层

Intermediate layers provide a set of abstractions for applications and media

中间层为应用程序和媒体提供了一组抽象
New applications or media only need implementation for intermediate layer's interface

新的应用程序或媒体只需要实现中间层的接口

Layering in the internet - OSI Model

互联网分层——OSI 模型

Application 应用层

Provides functions to users 为用户提供功能

Presentation 表示层

Converts different representations 转换不同的表示形式

Session 会话层

Manages task dialogs 管理任务对话框

Transport 传输层

Provides end-to-end delivery 提供端到端交付

Network 网络层

Sends packets over multiple links 通过多个链路发送数据包

Data Link 数据链路层

Sends frames of information 发送信息帧

Physical 物理层

Sends bits as signals 以信号形式发送比特

Network Socket-based Communication

基于网络套接字的通信

Socket

套接字

A software structure within a network node of a computer network that serves as an endpoint for sending and receiving data across the network.

计算机网络的网络节点内的软件结构，用作通过网络发送和接收数据的端点。

A Socket is combination of an lP Address and a Port Number.

Socket 是 IP 地址和端口号的组合。

// server side
struct sockaddr_in saddr, caddr;
int sockfd, clen, isock;
unsigned short port = 80;

if ((sockfd=Socket(AF INET,SOCK STREAM,0) < 0) {
    printf("Error creating socket\n");
    ...
}
    
memset(&saddr, '\0', sizeof(saddr));  // zero structure out
saddr.sin family = AF INET;  // match the socket() call
saddr.sin addr.s addr = htonl(INADDR ANY); // bind to any local addresss
addr.sin port = htons(port);  // specify port to listen on

if((bind(sockfd, (struct sockaddr *)&saddr, sizeof(saddr)) < 8) { // bind
    printf("Error binding\n");
    ...
}

if(listen(sockfd, 5) < 0) {  // listen for incoming connections
    printf("Error listening\n");
    ...
}

clen=sizeof(caddr)
if((isock=accept(sockfd, (struct sockaddr *)&caddr, &clen)) < 8) { // accept one
    printf("Error accepting\n");
    ...
}

Setup Socket

安装套接字

Both client and server need to setup the socket

客户端和服务器都需要设置套接字

Function
int socket(int domain, int type, int protocol)
- domain
  - AF_INET -- IPV4
  - AF_INET6 -- IPV6
    
    ...
  - type
    - SOCK_STREAM -- TCP
    - SOCK_DGRAM -- UDP
      
      ...
  - protocol
    - 0
- E0
  int sockfd = socket(AF_INET, SOCK_STREAM, 0)

Transmission Control Protocol (TCP)

传输控制协议（TCP）

Connection-Oriented

面向连接
Reliable and Ordered

可靠且有序
Flow Control

流量控制
Congestion Control

拥堵控制

TCP Retransmission Mechanisms

TCP 重传机制

Duplicate ACKs

重复 ACK

Senders sends packets and seqnos

发送方发送数据包和序列号
- 1,2, 3, 4, 5, 6, 7, 8
Assume 5th packet (seqno 5) is lost, Streamof ACKs will be

假设第 5 个数据包（序列号 5）丢失，则 ACK 流将
- 1, 2, 3, 4, 4, 4, 4

TCP congestion control

TCP 拥塞控制

The congestion window (CWND) of TCP starts to grow from a small value

TCP 的拥塞窗口（CWND）从一个小值开始增长
When congestion occurs, packet loss or timeout, CWND will be reduced according to a certain algorithm

当发生拥塞、丢包或超时时，CWND 将根据特定算法减少
This leads to high delay and cause delay jitter

这会导致高延迟并导致延迟抖动

As the main transmission protocol on the lnternet, TCP congestion control is necessary, otherwise it will cause congestion collapse. TCP congestion control is the main congestion control measure on the Internet, and it is also the main cause of TCP performance problems.

作为互联网上的主要传输协议，TCP 拥塞控制是必要的，否则会导致拥塞崩溃。TCP 拥塞控制是互联网上主要的拥塞控制措施，也是 TCP 性能问题的主要原因。

User Datagram Protocol (UDP)

用户数据报协议（UDP）

David P. Reed

He was involved in the early developmentof TCP/lP, and was the designer of the User Datagram Protocol (UDP), thoughhe finds this title "a little embarrassing".

他参与了 TCP/lP 的早期开发，是用户数据报协议（UDP）的设计者，尽管他觉得这个头衔“有点尴尬”。

He was also one of the authors of theoriginal paper about the end-to-end principle, End-to-end arguments insystem design, published in 1984.

他也是 1984 年发表的关于端到端原理系统设计中的端到端论证的原始论文的作者之一。

UDP Features

UDP 功能

UDP (User Datagram Protocol)

UDP（用户数据报协议）

Connectionless

无连接
UnReliable and Unordered

不可靠和无序
NO Flow Control

无流量控制
NO Congerstion Control

无堵塞控制

Network Protocols Usage in Game

Game suitable for use

TCP

炉石
UDP

守望先锋、CSGO

	TCP	UDP
Structure 结构	Segments 部分	Datagrams 数据报
Connection Model 连接模型	Connection-Oriented 面向连接的	Connectionless 无连接的
Speed 速度	Slow 慢	Fast 快
Reliability 可靠性	Reliable 可靠	Unreliable 不可靠
Header 标头	20 Bytes	8 Bytes
Data Transfer 数据传输	Ordered 有序的	Unordered 无序的
Control 控制	Flow Control 流量控制	NO Flow Control 无流量控制

Reliable UDP

可靠 UDP，在 UDP 的基础上加以改进。

TCP is Not Time Critical

TCP 不重视及时性

TCP is the complex and heavyweight protocol. lt provides reliable delivery and advanced features, but it has more overhead.

TCP 是复杂而重量级的协议。它提供可靠的交付和高级功能，但开销更大。
TCP is a fair, traffic oriented protocol designed to improve bandwidth utilization. But it's not designed for speed.

TCP 是一种公平的、面向流量的协议，旨在提高带宽利用率。但它不是为速度而设计的。
So Why TCP is slow?

为什么 TCP 很慢？

UDP is Fast but Unreliable

UDP 速度快但不可靠

UDP is lightweight and fast but unreliable, packet loss and disorder will occur.

UDP 是轻量级和快速的，但不可靠，会发生数据包丢失和混乱。
How to achieve reliable and real-time communication?

如何实现可靠和实时的通信？

Why We Need to Customize Protocol

为什么我们需要定制协议

Game Server

游戏服务器
- Keep-alived connection (TCP)
  
  保持有效连接（TCP）
- Need keep logic consistency in "order" (TCP)
  
  需要保持“顺序”（TCP）中的逻辑一致性
- High responsive & low latency (UDP)
  
  高响应和低延迟（UDP）
- Broadcast commonly used (UDP)
  
  常用广播（UDP）
Web Server

Web 服务器
- Handles the HTTP protocol
  
  处理 HTTP 协议
- Delivers static web content e.g., HTML pages, files, images, video.
  
  提供静态网络内容，例如 HTML 页面、文件、图像、视频。

Acknowledgement & Sequence Number

确认和序列号

Positive acknowledgment (ACK) is a signal that is passed between communicating processes, computers, or devices to signify acknowledgment, or receipt of message

肯定应答（ACK）是在通信进程、计算机或设备之间传递的信号，表示确认或收到消息
Negative ACK (NACK or NAK) is a signal that is sent to reject a previously receivedmessage or to indicate some kind of error

否定ACK（NACK 或 NAK）是一种信号，用于拒绝之前接收到的消息或指示某种错误
Sequence number (SEQ) is a counter used to keep track of every byte sent outward bya host

序列号（SEQ）是一个计数器，用于跟踪主机向外发送的每个字节
Timeouts specified periods of time allowed to elapse before an acknowledgment is tobe received

超时在收到确认之前允许经过的特定时间段

Automatic Repeat Request (ARQ)

自动重复请求（ARQ）

An error-control method for data transmission that uses ACK and timeouts to achieve reliable data transmission over an unreliable communication channel.

一种用于数据传输的错误控制方法，该方法使用 ACK 和超时来在不可靠的通信信道上实现可靠的数据传输。

If the sender does not receive an acknowledgment before the timeout, it re-transmits thepacket until it receives an acknowledgment or exceeds a predefined number of retransmissions.

如果发送方在超时之前没有收到确认，它会重新发送数据包，直到收到确认或超过预定义的重新传输次数。

Sliding window protocol

滑动窗口协议
- Stop-and-Wait ARQ
  
  停止并等待 ARQ
- Go-Back-N ARQ
- Selective Repeat ARQ
  
  选择性重复 ARQ

Sliding Window Protocol

滑动窗口协议

Send mutilple frames at a time, number of frames to be sent is based on Window size

一次发送多个帧，要发送的帧数基于窗口大小
Each frame is numbered by Sequence number

每帧按序列号编号
When the frame at the front of the window is received, the window slides

当收到窗户前部的框架时，窗户会滑动

Stop-and-Wait ARQ

停止并等待 ARQ

Windows size = 1

窗口大小 = 1
After transmitting one frame, the sender waits for an ACK before transmitting the next frame

发送一帧后，发送方在发送下一帧之前等待 ACK
If the ACK does not arrive after a certain time, the sender times out and retransmits the original frame

如果 ACK 在特定时间后未到达，则发送方超时并重新传输原始帧
Poor utilization of bandwidth, poor performance

带宽利用率低，性能差

Go-Back-N ARQ

N is Sender's Windows Size

N 是发件人的窗口大小
The Receiver only sends cumulative ACK

接收器仅发送累积 ACK
If an ACK is not received within an agreed-upon time period, all frames in the current window are transmitted

如果在商定的时间段内未收到 ACK，则传输当前窗口中的所有帧

Selective Repeat ARQ

选择性重复 ARQ

In Selective Repeat ARQ, only the damaged or lost frames are retransmitted

在选择性重复 ARQ 中，只有损坏或丢失的帧会被重新传输
The receiver sends the ack of each frame, and the sender maintains the timeout time of each frame

接收方发送每帧的 ack，发送方保持每帧的超时时间
When receiver receive damaged packet, it willsend a NACK, The sender will send/retransmitframe for which NACK is received

当接收方收到损坏的数据包时，它将发送一个 NACK。发送方将发送/重新发送收到 NACK 的帧

Make UDP Reliable in Packet Loss Scenario

在丢包情况下使 UDP 可靠

With the increase of packet loss rate and delay, the reliable UDP can not meet the transmission requirements gradually.eg. lf packet loss rate increase to 20%, use reliable UDP is still with high delay.

随着丢包率和延迟的增加，可靠的 UDP 逐渐无法满足传输要求。例如，如果丢包率增加到 20%，使用可靠的 UDP 仍然具有高延迟。

Forward Error Correction (FEC)

前向纠错（FEC）

The transmission of enough additiona, redundant information with the primary data stream to reconstruct lost lP packets up to a certain extent

与主数据流一起传输足够的附加、冗余信息，以在一定程度上重建丢失的 IP 数据包

XOR-FEC

A	B	A XOR B
0	0	0
0	1	1
1	0	1
1	1	0

Law 律	Desc 描述
Law of identity 同一律	A xor 0 = A
Zeroing law 零律	A xor A = 0
Commutative law 交换律	A xor B = B xor A
Associative law 结合律	A xor (B xor C) = (A xor B) xor C

C = A xor B

A = A xor (B xor B) = (A xor B) xor B = C xor B

B = (A xor A) xor B = A xor C

There are four packets A, B, C, D
- Let E = XOR (A, B, C, D)
- A = XOR (B, C, D, E)
- B = XOR (A, C, D, E)
- C = XOR (A, B, D, E)
- D = XOR (A, B, C, E)

If any packet is lost, it can be recovered with the other four packets.

如果任何一个数据包丢失，可以用其他四个数据包恢复。

Only one packet can be lost in continuous data. lf $A$ and $B$ are lost at the same time, the algorithm cannot recover.

连续数据中只能丢失一个数据包。如果 $A$ 和 $B$ 同时丢失，则算法无法恢复。

Reed-Solomon Codes

里德-所罗门码

There are $N$ valid data, and $M$ FEC data are expected to be generated

有 $N$ 个有效数据，预计将生成 $M$ 个 FEC 数据

Form $N$ valid data into a unit vector $D$

将 $N$ 个有效数据转换为单位向量 $D$
Generate a transformation matrix $B$: it is composed of a $N$-order identity matrix and a $N * M$ Vandemode matrix (The matrix composed of any n rows of
matrix $B$ is reversible)

生成一个变换矩阵 $B$：它由一个 $N$ 阶恒等式矩阵和一个 $N * M$ Vandemode 矩阵组成（该矩阵由以下任意 $N$ 行组成矩阵 $B$ 是可逆的）
The matrix $G$ obtained by multiplying the matrix $B$ and Vector $D$ contains $M$ redundant FEC data

通过将矩阵带向量 $D$ 相乘获得的矩阵 $G$ 包含 $M$ 个冗余 FEC 数据

Assume $D1$, $D4$, $C2$ are lost

假设 $D1$、$D4$、$C2$ 丢失

The $B$ matrix also needs to delete the corresponding $M$ rows to obtain a deformation matrix of $B'$

$B$ 矩阵还需要删除相应的 $M$ 行，以获得 $B'$ 的变形矩阵

Inverse matrix $B'$ get $B^{-1}$

逆矩阵 $B'$ 得到 $B^{-1}$
Multiply $B'^{-1}$ on both sides to recover the original data

将两侧的 $B'^{-1}$ 相乘以恢复原始数据

Customize Your UDP based on ARQ and FEC

基于 ARQ 和 FEC 自定义 UDP

Reliability

可靠性

Use Selective Repeat ARQ

使用选择性重复 ARQ

Hybrid ARQ and FEC

混合 ARQ 和 FEC

Before ARQ, FEC is used for error correction

在 ARQ 之前，FEC 用于纠错

Real-time

实时

Smaller RTO growth

RTO 增长较小
No congestion control

无拥堵控制
Fast retransmission mechanism

快速重传机制
No delay ACK

无延迟确认

Flexibility

灵活性

Design protocol for speed

速度设计协议
Support both reliable and unreliable transmission

支持可靠和不可靠的传输

Clock Synchronization

时钟同步

RTT

Round-Trip Time

往返时间

Send/Recv delay

发送/接收延迟
Propagation delay

传播延迟
Response time of the origin server

源服务器的响应时间

RTT vs. Ping

Ping tests are usually performed within a transport protocol that uses lCMP packets

Ping 测试通常在使用 lCMP 数据包的传输协议中执行
RTT is measured at the application layer

RTT 在应用层进行测量

RTT vs. Latency

Latency is the time required for a data packet to travel from the sending endpoint to the receivingend point (only one trip)

延迟是数据包从发送端点传输到接收点所需的时间（仅一次行程）

Network Time Protocol (NTP)

网络时间协议（NTP）

Network lime Protocol is an internet protocol used tosynchronize with computer clock time sources in a network.

网络时间协议是一种用于与网络中的计算机时钟时间源同步的互联网协议。

Reference clock

参考时钟
- GPS clock or radio ransmiiting station
  
  GPS 时钟或无线电传输站
- Amazinglly precise timekeeping devices such as atomic clocks
  
  令人惊叹的精确计时设备，如原子钟
- Not connected to the internet
  
  未连接到互联网
- Send their time through radio or optical fiber
  
  通过无线电或光纤发送他们的时间

Time Server Stratums

时间服务器层

Stratum Values

Degrees of separation from the reference clock

与参考时钟的分离程度
Reference clock has stratum value of 0

参考时钟的层值为 0
Servers with stratum value 1 is called primary time servers

层值为 1 的服务器称为主时间服务器
If a device's stratum value is over 15, its time is not trustworthy

如果设备的层值超过 15，则其时间不可信
Device will choose server with less stratum value automatically when correcting time

设备在校正时间时会自动选择层值较小的服务器

NTP Algorithm

NTP 算法

Use NTP is quite simple, just like this

使用 NTP 非常简单，就像这样

Client ask time server for time

客户端向时间服务器请求时间
Server receives the request and reply

服务器接收请求和回复
Client receives the reply

客户端收到回复

But we have to do something with the Delay!

但我们必须对延迟做点什么！

We recordes 4 timestamps as $t^c_0, t^s_1, t^s_2, t^c_3$

$\mathrm{Round\ Trip\ Delay}=(t^c_3-t^c_0)-(t^s_2-t^s_1)$

$\mathrm{Offset}=\frac{(t^s_1-t^c_0+t^s_2-t^c_3)}{2}$

The implicit assumption that the one way delay is statistically half the roundtrip delay

单向延迟在统计上是往返延迟的一半的隐含假设

Local-clock correction is computed from the offset data by.

根据偏移数据计算本地时钟校正。

$t^c_3 + \mathrm{Offset}$

*The delay and clock-offset samples obtained can be filtered using maximum-likelihood technigues

*获得的延迟和时钟偏移样本可以使用最大似然技术进行滤波

Let's take an example:

让我们举一个例子：

It's 17:01:00 on the client

$t^c_0$ is 17:01:00
$t^s_1$ is 17:01:32
$t^s_2$ is 17:01:33
$t^c_3$ is 17:01:05

It's 17:01:30 on the server

(17:01:30)

(17:01:32)

(17:01:33)

(17:01:35)

$\mathrm{Round\ Trip\ Delay}=(t^c_3-t^c_0)-(t^s_2-t^s_1)$

$\mathrm{Offset}=\frac{(t^s_1-t^c_0+t^s_2-t^c_3)}{2}$

Round Trip Delay = (05 - 00) - (33 - 32) = 4s

Offset = (32 - 00 + 33 - 05) / 2 = 30s

So at $t^c_3$ client's time is corrected to:

$t^c_3 + \mathrm{Offset}= 17:01:35$

Stream-Based Time Synchronization with Elimination of Higher Order Modes

基于流的时间同步，消除高阶模式

1. Client stamps current local time on a "time request" packet and sends to server

客户端在“时间请求”数据包上标记当前本地时间并发送到服务器

2. Upon receipt by server, server stamps server-time and returns

服务器收到后，服务器标记服务器时间并返回

3. Upon receipt by client, a time delta is calculated by delta = (current Time-sent Time) / 2

客户端收到后，通过 delta=（当前发送时间）/2 计算时间增量

So far this algorithm is very like to NTP

到目前为止，该算法与 NTP 非常相似

4. The first result should immediately be used toupdate the clock

第一个结果应立即用于更新时钟

5. The client repeats Steps 1-3 (NTP-like process), five or more times

客户端重复步骤1-3（类似 NTP 的过程）五次或更多次

6. The results of the packet receipts are accumulatedand sorted in ascending order by latency

数据包接收的结果按延迟升序累积和排序

7. All samples above that are approximately 1.5 times the median are discarded, and the remaining samples are averaged using an arithmetic mean

丢弃上述约为中位数 1.5 倍的所有样本，并使用算术平均值对剩余样本进行平均

Remote Procedure Call (RPc)

远程过程调用（RPC）

Socket Programming: Still not Great

Socket 编程：仍然不太好

Lots for the programmer to deal with every time

程序员每次都要处理很多事情
- How to separate different requests on the same connection?
  
  如何在同一连接上分离不同的请求？
- How to write bytes to the network/read bytes from the network?
  
  如何向网络写入字节/从网络读取字节？
- What if Host A's process is written in Go and Host B's process is in C++?
  
  如果主机 A 的进程是用 Go 编写的，而主机 B 的进程是 C++ 编写的呢？
- What to do with those bytes?
  
  如何处理这些字节？
Still pretty painful... have to worry a lot about the network

仍然很痛苦…不得不非常担心网络
- Have you received the message?
  
  你收到消息了吗？

Communication with Messages

消息通信

Communication Way

沟通方式

Initially, people "hand-coded" messages to send requests and responses

最初，人们“手工编码”消息以发送请求和响应

Message is a stream of bytes-"op codes" and operands

消息是字节流-“操作码”和操作数

Lots of drawbacks

很多缺点

Need to worry about message format

需要担心消息格式
Have to pack and unpack data from messages

必须从消息中打包和解包数据
Servers have to decode messages and dispatch them to handlers

服务器必须解码消息并将其分派给处理程序
Messages are often asynchronous

消息通常是异步的
- After sending one, what do you do until the response comes back?
  
  发送一个后，在收到回复之前，你会做什么？
Messages aren't a natural programming mode

消息不是一种自然的编程模式

struct foomsg {
    uint32 _t len;
};

void send_foo(char *contents) {
    int msglen = sizeof(struct foomsg) + strlen(contents);
    char buf = malloc(msglen);
    struct foomsg *fm = (struct foomsg *)buf;
    fm->len = htonl(strlen(contents));
    memcpy(buf + sizeof(struct foomsg), contents, strlen(contents));
    write(outsock,buf, msglen);
}

More Challenges on Logic Communication

逻辑通信面临更多挑战

For a remote procedure call, a remote machine may:

对于远程过程调用，远程机器可以：
- Run process written in a different language
  
  运行用其他语言编写的进程
- Represent data types using different sizes
  
  使用不同大小表示数据类型
- Use a different byte ordering (endianness)
  
  使用不同的字节顺序（字节序）
- Represent floating point numbers differently
  
  以不同的方式表示浮点数
- Have different data alignment requirements
  
  有不同的数据对齐要求
  
  e.g., 4-byte type begins only on 4-byte memory boundary
  
  例如，4 字节类型仅在 4 字节内存边界上开始

Remote Procedure Call (RPC)

远程过程调用（RPC）

RPC is a reguest-response protocol. An RPC is initiated by the cient, which sends a reguest message to a known remote server to execute a specified procedure with supplied parameters

RPC 是一种规范的响应协议。RPC 由 cient 发起，它向已知的远程服务器发送一条规则消息，以使用提供的参数执行指定的过程
Goals

目标
- Ease of programming
  
  易于编程
- Hide complexity
  
  隐藏复杂性
- Familiar model for programmers (just make a function call)
  
  程序员熟悉的模型（只需进行函数调用）

RPC example

Go language

Output

Hello World

Why RPC?

为什么选择 RPC？

Goal: Easy-to-program network communication that makes client-server communication transparent

目标：易于编程的网络通信，使客户端-服务器通信透明
- Retains the "feel" of writing centralized code
  
  保留编写集中式代码的“感觉”
- Programmers needn't think about the network
  
  程序员不必考虑网络
- Make communication appear like a local procedure call
  
  使通信看起来像本地过程调用
Don't need to worry about serialization/deserialization for network

无需担心网络的序列化/反序列化
Don't need to worry about complexities of network

无需担心网络的复杂性

Interface Definition Language

接口定义语言

A server defines the service interface using an interface definition language (lDL)

服务器使用接口定义语言（lDL）定义服务接口

The lDL specifies the names, parameters, and types for allclient-callable server procedures

lDL 指定了所有客户端可调用服务器过程的名称、参数和类型
- example: ASN.1 in the OSI reference model
  
  示例：OSI 参考模型中的 ASN.1
- example: Protobuf (Google's data interchange format)
  
  示例：Protobuf（谷歌的数据交换格式）
- ...

//polyline.proto
syntax ="proto2";

message Point {
    required int32 x = 1;
	required int32 y = 2;
    optional string label = 3;
}

message Line {
    required Point start = 1;
    required Point end = 2;
    optional string label = 3;
}

message Polyline {
    repeated Point point = 1;
    optional string label = 2;
}

RPC Stubs

A client-side stub is a procedure that looks to the client as if it were a callable server procedure

客户端存根是一个过程，它在客户端看起来就像是一个可调用的服务器过程
- The client program thinks it's invoking the server butit's calling into the client-side stub
  
  客户端程序认为它正在调用服务器，但它正在调用客户端存根
Aserver-side stub looks like a caller to the server

服务器端存根看起来像是服务器的调用者
- The server program thinks it's called by the client but it's really called by the server-side stub
  
  服务器程序认为它是由客户端调用的，但实际上是由服务器端存根调用的
The stubs send messages to each other to make the RPC happen transparently

存根相互发送消息，使 RPC 透明地发生

Stub compiler

A "stub compiler" reads the IDL declarations and produces two stub procedures for each server procedure

“存根编译器”读取 IDL 声明，并为每个服务器过程生成两个存根过程
- The server programmer implements the service's procedures and links them with the server-side stubs
  
  服务器程序员实现服务的过程，并将其与服务器端存根链接起来
- The client programmer implements the client program and links it with the client-side stubs
  
  客户端程序员实现客户端程序并将其与客户端存根链接
- The stubs manage all of the details of remote communication between client and server
  
  存根管理客户端和服务器之间远程通信的所有细节

Real RPC Package vourney

真实 RPC 包凭证

Network Topology

网络拓扑

Original Peer-to-Peer (P2P)

原始点对点（P2P）

Each client broadcasts game event to the al others

每个客户端向其他客户端广播游戏事件
Robustness

稳健性
Cheating is much easier

作弊要容易得多
Synchronization is required among all nodesto maintain the consistency of the distributed game state

所有节点之间需要同步，以保持分布式游戏状态的一致性

P2P with Host Server

P2P 与主机服务器（如 CS）

A player can act as "server", known as host

玩家可以充当“服务器”，即主机
If host disconnected, the game may end

如果主机断开连接，游戏可能会结束
The host need to handle game actor that cannot be controled by players, such as bot

主机需要处理玩家无法控制的游戏角色，如机器人

P2P Games

P2P 游戏

No rely on server

不依赖服务器
Used in Lan commonly

常用于局域网
The “Host”is basically in control of the sessions

“主持人”基本上控制着会话
A limited number of players at once

一次玩家数量有限

Dedicated Server

专用服务器

Authority

权威机构
Simulate game world

模拟游戏世界
Dispatch data to players

向玩家发送数据
High performance requirements

高性能要求

P2P vs Dedicated Server

P2P 与专用服务器

	P2P	Dedicated Server
Pros	1. Robustness 稳健性 2. Removes the "server issues" problem in multiplayer sessions. 消除多人游戏会话中的“服务器问题”问题 3. No extra cost on server 服务器无需额外付费	1. Easy to maintain as well as cheating avoidance 易于维护，避免作弊 2. Can handle massive game world 可以处理大型游戏世界 3. Responsiveness of the game is not relay on the network conditions of each individual client 游戏的响应性并不取决于每个客户端的网络状况
Cons	1. Cheating is much easier 作弊要容易得多 2. Every player needs a decent network connection for game to function properly 每个玩家都需要一个像样的网络连接，游戏才能正常运行 3. Can only handle a limited number of players 只能处理有限数量的玩家	1. High cost on server 服务器成本高 2. Much more work on server side program 在服务器端程序上还有更多的工作要做 3. Single point of failure 单点故障

When RTT is too high

RTT 过高时

When players are in different countries, far away, or when the network environment is complex

当玩家身处不同的国家、遥远的地方，或者网络环境复杂时
Use dedicated line and edge gateway to reduce latency

使用专用线路和边缘网关来减少延迟

Game Synchronization

游戏同步

Single-Player Gaming

Game Tick

Player inputs

玩家输入
Convert to game commands

转换为游戏命令
Game logic

游戏逻辑
Game render

游戏渲染

For Player

对于玩家

Player inputs

玩家输入
Consistency in each other

彼此的一致性

Online Gaming

在线游戏

For Player

对于玩家

Player inputs

玩家输入
Consistency in each other

彼此的一致性

How to play together at different terminals?

如何在不同的终端上一起玩？

Game commands

游戏命令
Game Logic

游戏逻辑

Game Synchronization

游戏同步

To answer the demand for responsive strategies, the synchronization rule is designed to solve the delay and consistency of all destination.

为了满足对响应策略的需求，设计了同步规则来解决所有目的地的延迟和一致性问题。

Synchronization Methods

同步方法

Snapshot 快照	Lockstep 锁步	State Synchornization 状态同步
Quake	Honor of Kings 王者荣耀	Counter Strike 反恐精英

Snapshot Synchronization

快照同步

Client sends inputs to server

客户端向服务器发送输入
Server simulates the game world

服务器模拟游戏世界
Generates whole game statesnapshots

生成整个游戏状态快照
Sends them down to clients

将它们发送给客户
Client updates the display according to the snapshot

客户端根据快照更新显示

Snapshot Synchronization - Jitter and Hitches

快照同步-抖动和挂接

Server tick rate is limited

服务器时间速率有限
Performance

性能
Bandwidth

带宽

Snapshot Interpolation

快照插值

Not rendering immediately after snapshot recevied

收到快照后不立即渲染
Keep an interpolation buffer

保留插值缓冲区
Interpolation between the two delayed snapshots

两个延迟快照之间的插值

Delta Compression

德尔塔压缩

Only sync snapshot delta to client

仅将快照增量同步到客户端
Example Quake3

Snapshot Synchorization

快照同步

60HZ
Max Bandwidth 4.5 mbps

最大带宽 4.5mbps

Delta Compression

德尔塔压缩

60HZ
Max Bandwidth 676 kbps

最大带宽 676 kbps

Synchronizing Snapshot

正在同步快照

Client performance is wasted

客户端性能被浪费
High server pressure

服务器压力高
High data volume and high bandwidth requirements

高数据量和高带宽要求
As games get more complex, snapshots get bigger

随着游戏变得越来越复杂，快照也越来越大

Lockstep Synchronization

锁步同步

Lockstep Origin

锁步原点

Lockstep synchronization, used in military simulation, is by far the simplest techique to ensure consistency

用于军事模拟的锁步同步是迄今为止确保一致性的最简单技术。

Same Result

相同的结果
Same time

同一时间
Same action

同样的行动

No member is allowed to advance its simulation clock until all others members have acknowledged that they are done.

在所有其他成员确认完成之前，任何成员都不允许提前其模拟时钟。

In particular, it is clear that a totally ordered delivery is a sufficient condition to ensure game state consistency across different nodes, as it guarantees that all generated events are reliably delivered according to the same unique order.

特别是，很明显，完全有序的交付是确保不同节点之间游戏状态一致性的充分条件，因为它保证了所有生成的事件都按照相同的唯一顺序可靠地交付。

Lockstep in Online Game

网络游戏中的锁步

Lockstep Principle.

锁步原理。

First Game Used Lockstep

第一个游戏使用锁步

The network synchronization methodof DOOM (1994) was pointed out in a 2006 paper

2006 年的一篇论文指出了 DOOM（1994）的网络同步方法
Lockstep is not mentioned in the paper, but it is now generally accepted that Doom (1994) was the first multiplayer FPS online game to use this type of synchronization

论文中没有提到锁步，但现在人们普遍认为《毁灭战士》（1994）是第一款使用这种同步的多人 FPS 在线游戏
It uses P2P architecture

它使用 P2P 架构

Lockstep initialization

锁步初始化

加载…

Ensure that the initial data of each client is deterministic

确保每个客户端的初始数据是确定的
- Game model
  
  游戏模型
- Static data
  
  静态数据
- ...
Synchronize clock

同步时钟

Deterministic Lockstep

确定性锁步

Client sends inputs to Server

客户端向服务器发送输入
Server receives and sorts

服务器接收和排序
Wait for input from all clients before forwarding

转发前等待所有客户端的输入
After receiving data from the server, the client executes the game logic

在从服务器接收到数据后，客户端执行游戏逻辑

If Player B's message B2 arrives later? (The dotted line B2 in the figure)

如果玩家 B 的消息 B2 稍后到达？（图中虚线 B2）

Disadvantages

缺点
- Game progress depends on slowest player
  
  游戏进度取决于最慢的玩家
- The delay of the game is notfixed, the experience is not good
  
  游戏的延迟不是固定的，体验不好
- All the players will wait if a player offline
  
  如果有玩家离线，所有玩家都会等待

Player Offline in Deterministic Lockstep

确定性锁步中的玩家离线

Waiting for players...

正在等待玩家...

Bucket Synchronization

Bucket 同步

Bucket: a fixed time period

Bucket：固定时间段
Each bucket

每个水桶
- Colect all instructions
  
  整理所有说明
- Broadcast to all players
  
  向所有玩家广播
There is no need to wait for all players' commands to be received before forwarding

转发前无需等待收到所有玩家的命令

A Good Trade-off between Consistency and Interactivity Maintenance

一致性和交互性维护之间的良好权衡

So we need to find a basis balance between them.

因此，我们需要在它们之间找到一个基本的平衡点。

The threshold:

阈值：

As soon as the measured interactivity degree decreases below a given thresholdtake some procedure skips processing obsolete game events withthe aim of bringing back a satisfactory interactivity level.

一旦测量的交互性降低到给定的阈值以下，一些程序就会跳过处理过时的游戏事件，以恢复令人满意的交互性水平。

Deterministic Difficulties

确定性难题

Deterministic

确定性
- The same input sequence need to produce the same game state on all machines
  
  相同的输入序列需要在所有机器上产生相同的游戏状态
Deterministic is Hard

确定性很难
- Floating point
  
  浮点型
- Random number
  
  随机数
- Containers and algorithms (sort, add, remove, etc.)
  
  容器和算法（排序、添加、删除等）
- Math tools (vectors, quaternions, etc)
  
  数学工具（向量、四元数等）
- Physics simulation (very difficult)
  
  物理模拟（非常困难）
- Code logic execution order
  
  代码逻辑执行顺序

Floating Point Numbers

浮点数

Because of the computer binary, Thesenumbers can be accurately represented

由于计算机二进制，这些数字可以被精确地表示出来
- 0.5 = 1/2
- 0.25 = 1/4
- 0.75 = 1/2 + 1/4
- 0.875 = 1/2 + 1/4 + 1/8
Such numbers can only be approximated

这些数字只能近似表示
- 2/3 ≈ 0.66..7

Floating point numbers must comply with the lEEE 754 standard

浮点数必须符合 lEEE 754 标准

Floating point numbers conform to the lEEE 754 standard

浮点数符合 lEEE 754 标准
But different platforms may have different behavior

但不同的平台可能有不同的行为

Floating Point Hardware & OS Behaviour

浮点硬件和操作系统行为

Intel / Amd
PS / Xbox
Windows / linux
Android / IOS
...

Floating Point Compilers Behaviour

浮点编译器行为

Math Library(sin、cosin、tan、exp、 pow ...)

数学库（sin、cosin、tan、exp、pow…）
Third party components

第三方组件
Different platforms

不同的平台
Different versions

不同版本
Different languages

不同的语言
...

Idea: Avoid problems on the precision boundary, customize the precision

理念：避免精度边界问题，定制精度

Fixed-point math library

定点数学库
Look-up table (trigonometric functions, etc.)

查找表（三角函数等）
Amplification and truncation

放大和截断

Simple method

简单的方法

Multiply by 1000, then divide by 1000, there is an overflow risk

乘以 1000，然后除以 1000，存在溢出风险
The numerator and denominator are represented by fixed-point numbers (2/3)

分子和分母由定点数（2/3）表示
...

One Solution: Fixed point math

一种解决方案：定点数

A fixed-point number can be split into three parts

一个定点数可以分成三部分

An optional sign bit

可选符号位
An integer

一个整数
A fractional part

分数部分
Need to implement addition, subtraction, multiplication and division etd.

需要实现加、减、乘、除等运算
Implement class, class methods

实现类、类方法
Performance needs to be considered

需要考虑性能

$V=(-1)^{b_{f+i}}(\sum_{n=0}^{i}2^{n}b_{n+f}+\sum_{m=1}^{f}2^{-m}b_{f-m})$

Random Number

随机数

Random problems in the game

游戏中的随机问题
- Trigger of random events, npc random birthplace
  
  随机事件触发，NPC 随机出生地
- A random attribute of the attack, e.g. critical strike chance
  
  攻击的随机属性，例如暴击机会
- ...
These logics are generally implemented with random numbers

这些逻辑通常用随机数实现
How to implement random logic that is completely consistent for multiple players

如何为多个玩家实现完全一致的随机逻辑

Random numbers are pseudorandom

随机数是伪随机的
Before the game starts, initialize the random number seed

在游戏开始之前，初始化随机数种子
For different players' clients, the number of random function calls is fixed, and the generated random numbers are the same

对于不同玩家的客户端，随机函数调用的数量是固定的，生成的随机数是相同的

int main() {
	std::default random engine e;
    std::uniform int distribution<int> u(0,100);
    e.seed(80);
    
	for(int i = 0; i < 20; i++) {
        std::cout << u(e) << std::endl;
    }
    
	return 0;
}

Deterministic Solution

确定性解

Fixed-point numbers represent floating-point numbers in critical game logic

定点数表示关键游戏逻辑中的浮点数
Deterministic random algorithm

确定性随机算法
Deterministic containers and algorithms (sort, add, remove, etc.)

确定性容器和算法（排序、添加、删除等）
Deterministic math tools (vectors, quaternions, etc.)

确定性数学工具（向量、四元数等）
Deterministic physics simulation (very difficult)

确定性物理模拟（非常困难）
Deterministic execution order

确定性执行顺序

Tracing and Debugging

跟踪和调试

Method of get checksum

获取校验和的方法

All data checksum

所有数据校验和
Key data checksum

关键数据校验和
Other methods

其他方法

Automatically locate BUG

自动定位BUG

Server compares different client's checksums

服务器比较不同客户端的校验和
Client uploads 50 frames of full logs

客户端上传 50 帧完整日志
Find inconsistencies in the compared logs

查找比较日志中的不一致之处

Lag and Delay

延迟和延误

Client send operation

客户端发送操作
Receive the operation of this frame from the server

从服务器接收此帧的操作
execute

执行

Lag: Network is unstable. lf you wait until you receivenew frame, there will be a lag

滞后：网络不稳定。如果你等到收到新帧，就会产生延迟

Solution

解决方案

use buffer to cache frames

使用缓冲区缓存帧
- Large buffer, large delay
  
  缓冲区大，延迟大
- Small buffer, sensitive to lag
  
  小缓冲区，对延迟敏感

Separating Game Logic from Rendering

将游戏逻辑与渲染分离

Lag problem

滞后问题

Separation of logic and rendering

逻辑与渲染的分离
Local client-side interpolation smoothing

本地客户端插值平滑

Frame rate

帧率

The logical frame rate is generally 10~30 frames

逻辑帧率一般为 10~30 帧
The rendering frame rate is generally higher

渲染帧率通常较高

Advantage

优势

Different frequencies, independent operation

不同频率，独立运行
Rendering separation to avoid tearing and freezing

渲染分离以避免撕裂和冻结
Rendering freezes, does not affect the operation of logical frames

渲染冻结，不影响逻辑帧的操作
Servers can run logic frames to solve some cheating problems

服务器可以运行逻辑框架来解决一些作弊问题
If the server runs logical frames, it can save key frame snapshots to speed up reconnection

如果服务器运行逻辑帧，它可以保存关键帧快照以加快重新连接

Reconnection Problem

重新连接问题

Offline

离线
Reconnect

重新连接
Catch up

赶上

Client Game State Snapshots

客户端游戏状态快照

Snapshots can be saved regularly on the local client and serialized to disk

快照可以定期保存在本地客户端上并序列化到磁盘
When reconnection occurs, restore thegame state from the disk serialized data

重新连接时，从磁盘序列化数据还原游戏状态
Server sends player commands after snapshot

服务器在快照后发送玩家命令
Accelerate to catch up with the game progress

加速以赶上游戏进度

Quick Catch Up

快速赶上

How to catch up?

如何迎头赶上

In the sample code, chasing 10 frames each time

在示例代码中，每次追踪 10 帧
If originally 10 frames per second, when chasing frames, it may run 100 frames per second

如果最初每秒 10 帧，在追逐帧时，它可能每秒运行 100 帧

float m_delta = 0;
float m_tick_delta = 100;
void CBattleLayer::update(float delta) {
    // do something
    m_delta += delta;
    int exec_count = 1;
    while(m_delta >= m_tick_delta) {
        m_delta -= m_tick_delta;
        // logic frame
        if(!logicUpdate(LOGIC_TIME)) {
            return;
        }
        // catch up 10 frames at a time
        if(exec_count++ >= 10){
            break;
        }
    }
	// do something
}

Server State Snapshot Optimization

服务器状态快照优化

The server runs logical frames and saves snapshots of keyframes

服务器运行逻辑帧并保存关键帧的快照
The server sends the snapshot, and the player commands after the snapshot

服务器发送快照，玩家在快照后发出命令
Accelerate to catch up with the game progress

加速以赶上游戏进度

Temporary Offline, No Crash

暂时离线，无崩溃

Client also keeps game state, keyframes, deterministic-timed frames

客户端还保留游戏状态、关键帧、确定性定时帧
After reconnecting, the server sends commands to the dropped player

重新连接后，服务器向被删除的播放器发送命令
Accelerate to catch up the game progress

加速以赶上游戏进度

Observing

观察

Watching other players playing the game

观看其他玩家玩游戏

Reconnecting and watching are essentially the same

重新连接和观看基本上是一样的
Watching is similar to reconnecting after a client crash

监视类似于在客户端崩溃后重新连接
Player action command, forwarded to the player watching the game

玩家动作命令，转发给观看游戏的玩家
Watching is usually delayed for a few minutes to prevent screen peeping

观看通常会延迟几分钟，以防止偷看屏幕

Replay

重播

Execute player commands in order which can speed up

按照可以加速的顺序执行玩家命令

Replay file

回放文件
- Save game commands for a game
  
  保存游戏命令
- Files take up little space
  
  文件占用的空间很小
How to implement go back?

如何实施回退？
- When the client executes the replay file, it adds a key frame snapshot, which can go back to the key frame moment
  
  当客户端执行回放文件时，它会添加一个关键帧快照，该快照可以返回到关键帧时刻
- The current version of Honor of Kings go back to the key frame before 60s
  
  当前版本的《王者荣耀》回退了 60s 的关键帧

Lockstep Cheating lssues

锁步作弊问题

Multiplayer-PVP

多人PVP

Game over

游戏结束
- The client uploads the key data checksum, the server verifys the game result
  
  客户端上传密钥数据校验和，服务器验证游戏结果
During the game

在游戏过程中
- Report the key data checksum
  
  报告密钥数据校验和
- Cheating players are kicked out, etc.
  
  作弊的玩家被踢出等等。

2 Players

2 名玩家

Server can not detect who is cheating using key data checksum

服务器无法使用密钥数据校验和检测谁在作弊
If the server is not verified, the cheating player will only affect one player in this case

如果服务器未经验证，作弊玩家在这种情况下只会影响一个玩家

Difficult to avoid third-party plug-in to access war-fog or other hidden data

难以避免第三方插件访问战争迷雾或其他隐藏数据
- Game logic is performed on the client side
  
  游戏逻辑在客户端执行
- Clients have all the game data
  
  客户端拥有所有游戏数据

Lockstep Summary

锁定步骤摘要

Advantages

优势

Low bandwidth, only sends commands

带宽低，只发送命令
High development efficiency, similar to single-player game development

开发效率高，类似单人游戏开发
Precise action/hit detection

精确的动作/命中检测
Easy to record games

易于录制游戏

Problems

问题

Maintain the consistency is difficult to achieve

保持一致性很难实现
Hard to solve the cheat plugin to unveil all game states

难以解决的作弊插件揭示所有游戏状态
Longer disconnection and reconnection time

更长的断开和重新连接时间
- Need more complex optimization
  
  需要更复杂的优化

State Synchronization

状态同步

Replication Protocol of Halo

Halo 的同步协议

State Data	Events	Control Data
Guaranteed eventual delivery of most current state 保证最终交付最新状态 - Object position 对象位置 - Object health 对象生命 - 150+ properties 150+ 个属性	Unreliable notifications of transient occurrences 不可靠的瞬态事件通知 - Please fire my weapon 请发射我的武器 - This weapon was fired 此武器已被发射 - Projectile detonated 弹丸被引爆 -More events 更多事件	High-frequency, the best-effort transmission of rapidly-updated data extracted from player control inputs 高频、尽最大努力传输从玩家控制输入中提取的快速更新数据 - Current analog stick values for all players 所有玩家的当前模拟棒值 - Current position of client's own biped 客户端自己的角色的当前位置 -More properties 更多属性

State

状态

The game state is necessary to represent the game world. e.g: HP, MP

游戏状态是代表游戏世界所必需的。例如：血量、法力

State Synchronization

状态同步

Server does not generate a single update for all clients. lt sends client a customized data packet

服务器不会为所有客户端生成单个更新。它向客户端发送自定义数据包
lf the game world is too complex, you can set an Area Of Interest (AOl) for reducing server overhead

如果游戏世界太复杂，您可以设置一个兴趣区域（AOl）来减少服务器开销

Server Authorizes the Game World

服务器授权游戏世界

Server

服务器

Game world is authorized

游戏世界已授权
Receive input and state from client

从客户端接收输入和状态
Run game logic

运行游戏逻辑
Send state

发送状态

Client

客户端

Receive data and simulate game world

接收数据并模拟游戏世界
Game play improvement

游戏玩法改进

Authorized and Replicated Clients

授权和同步客户端

Authorized (1P)

授权（1P）

Player's local game client

玩家的本地游戏客户端

Server

服务器

Authorized server

授权服务器

Replicated (3P)

已同步（3P）

Simulated character in other player's client

其他玩家客户端中的模拟角色

State Synchronization Example

状态同步示例

Player1 (Authorized)

玩家1（授权）

Fire

开火

Player2 (Replicated)

Player2（已同步）

See player1 open fire

看到玩家 1 开火

Player 1 presses an input on their local machine to fire

玩家 1 按下本地机器上的输入进行射击

Player1 (Authorized)

玩家 1（授权）

Fire

开火
Send to server

发送到服务器

Server

服务器

Player1 fire

玩家 1 开火
Send to each client

发送给每个客户端

Player2 (Replicated)

玩家 2（已同步）

Recieve packet

接收数据包
Player1 fire

Player 1 开火

Server

服务器

Tell each client to replicate the movement of Player 1's projectile.

告诉每个客户同步玩家 1 投射物的运动。

Server

Tell each client to destroy their copies of Player 1's projectile

告诉每个客户销毁他们的玩家 1 投射物副本
Tell all clients to response to damage of the projectile

告诉所有客户对弹丸的损坏做出反应

Dumb Client Problem

哑巴客户端问题

Clients can not to do anything until receive server state update

在收到服务器状态更新之前，客户端无法执行任何操作

How to see an immediate response?

如何看到即时响应？

Client-side prediction

客户端预测
Server reconciliation

服务器对账

Client-Side Prediction

客户端预测

Authorized client

授权客户

Press "→"

按“→”
Received server message

收到服务器消息
Start movement

开始移动

Overwatch - Client-side Prediction

守望先锋-客户端预测

RTT = 160ms

RTT = 160 毫秒
Half RTT = 80ms

半 RTT = 80 毫秒
Command frame = 16ms

命令帧 = 16ms

The client is always ahead of the server by half RTT and one buffered command frame

客户端始终领先于服务器一半的 RTT 和一个缓冲命令帧

Press key and response immediately

按键并立即响应

Server Reconciliation

服务器对账

Authorized client: Buffer

授权客户：Buffer

Record every state when do the client prediction

在进行客户端预测时记录每个状态
Compare with the past server data when it was received on client side

与客户端接收到的过去服务器数据进行比较

Ring buffer for states

状态环形缓冲区

Stores all of our states in the past several frames on client

将过去几帧中的所有状态存储在客户端上

Process

流程

If the client computed the same result as server, the client will continue on its merry way to simulate the next input

如果客户端计算出的结果与服务器相同，客户端将继续愉快地模拟下一个输入

Problem

问题

lf misprediction?

如果预测失误？

If blocked by an obstacle at the server

如果被服务器上的障碍物阻挡
Position is wrong! (in red)

位置不对！（红色）
Client must accept the new server

客户端必须接受新服务器
update

更新
Retrace all predicted movement starting from the new confirmed position

从新的确认位置开始回溯所有预测的运动

If the client and server disagree on the results, we've mispredicted

如果客户端和服务器对结果不一致，我们就预测错了
Have to reconcile

必须和解

Ring buffer for inputs

用于输入的环形缓冲区

Stores all of inputs we did in the past several frames on client.

在客户端上存储我们在过去几帧中所做的所有输入。

Process

流程

Overwrite the clients results withthe server's results

用服务器的结果覆盖客户端的结果
Replay all of your inputs to catchback up to what you believe now

重播您的所有输入，以回溯您现在所相信的内容

Server Reconciliation Example

服务器对账示例

Overwatch

守望先锋

We try to move

我们试图移动
The server said no

服务器拒绝了
We got yanked back down to where we were before and froze

我们被拉回到原来的位置，然后僵住了

Packet Loss

数据包丢失

Client input packages fail to reach the server

客户端输入包无法到达服务器
The server tries to keep tiny input buffer of unprocessed input

服务器试图保留未处理输入的微小输入缓冲区
lf the server run out of input buffer, server will duplicate your last input in a window

如果服务器的输入缓冲区用完，服务器将同步您在窗口中的最后一个输入
Push client sends missed inputs asap

推送客户端尽快发送丢失的输入

State Synchronization Vs. Lockstep Synchronization

	State Synchronization 状态同步	Lockstep Synchronization 锁步同步
Deterministic logic 确定性逻辑	Not necessary 不需要	Necessary 必要的
Response 响应	Better responsiveness 更好的响应能力	Poor responsiveness 响应性差
Network Traffic 网络流量	Usually high 通常高	Usually low 通常低
Development efficiency 开发效率	Much more complicated 要复杂得多	Easy to develop, difficult to Debug 易于开发，难以调试
Number of players 玩家数量	Few players 玩家很少	Support small and large numbers of players 支持小型和大型玩家
Cross-platform 跨平台	Relatively easy 相对容易	Relatively difficult 相对困难
Reconnect 重连	Relatively easy 相对容易	Relatively difficult 相对困难
Replay File size 重播文件大小	Big 大	Small 小
Cheat 欺骗	Relatively hard 相对困难	Relatively easy 相对容易

References

Network Protocols

Internet protocol suite:

https://en.wikipedia.org/wiki/lnternet_protocol_suite
OSI Model:

https://www.practicalnetworking.net/series/packet-traveling/osi-model
Automatic repeat request.

https://en.wikipedia.org/wiki/Automatic_repeat_request
Network Communication and Remote Procedure Calls (RPCs):

https://www.cs.princeton.edu/courses/archive/fall18/cos418/docs/L2-rpc.pdf
What are TCP and UDP?
https://www.gdyuniie.cn/showinfo-114-788-0.html

Network Synchronization

Synchronization lssues with Smart Solutions:

https://citeseerxist.psu.edu/viewdoc/download?doi=10.1.1.352.6405&rep=rep1&type=pdf
An Efficient Synchronization Mechanism for Mirrored Game Architectures:

https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.87.6043&rep=rep1&type=pdf
The Brave New World of Multiplayer Online Games - Synchronization lssues with Smart Solutions.Marco Roccetti, Stefano Ferretti, Claudio E. Palazzi:

https://citeseerxist.psu.edu/viewdoc/download?doi=10.1.1.352.6405&rep=rep1&type=pdf
Network Time Protocol(NTP): https://www.rfc-editor.org/rfc/rfc958.html
Algorithms for Synchronizing Network Clocks: https://datatracker.ietf.org/doc/html/rfc956
"Minimizing Latency in RealTime Strategy Games", Jim Greer, Zack Booth Simpson, GameProgamming Gems 3 chapter 5.1, 2001: https://larchive.org/details/game-programming-gems-3
JMP van Waveren, "The DOOM lll Network Architecture".2006:

https://mrelusive.com/publications/papers/The-DOOM-l-Network-Architecture.pdf
Christophe DlOT, Laurent GAUTlER, "A Distributed Architecture for Mult iplayer Interactive Applicationson the Internet", lEEE.1999:

https://ieeexplore.ieee.org/abstract/document/777437/
Mark Terrano, Paul Bettner "Network Programming in Age of Empires and Beyond". GDC 2001.

https://www.aamedevs.org/uploads/1500-archers-age-of-empires-network-programming.pdf
腾讯游戏, 腾讯游戏开发精粹.电子工业出版社,2019.9:

https://gameinstitute.gg.com/game-gems
Cocos, 帧同步游戏在技术层面的实现细节, 2021

https://zhuanlan.zhihu.com/p/408734657
IEEE 754:

https://en.wikipedia.org/wiki/IEEE_754
QUAKE 3 SOURCE CODE REVIEW - NETWORK MODEL. 2012:

https://fabiensanglard.net/quake3/network.php
David Aldridge, l Shot You First - Networking the Gameplay of HALO - REACH. GDC 2011:

https://www.gdcvault.com/play/1014345/1-Shot-You-First-Networking
Timothy Ford, "Overwatch Gameplay Architecture and Netcode". GDC 2017: https://www.gdcvault.com/play/1024001/-0verwatch-Gameplay-Architecture-and
Unreal engine Document:

https://docs.unrealengine.com/5.0/en-US/networking-overview-for-unreal-engine/
Gaffer on Games: https://gafferongames.com/#posts

第十九节：网络游戏的进阶架构

Online Gaming Architecture - Advanced Topics

Character Movement Replication

角色动作同步

Character Movement Replication

角色动作同步

From player 2's point of view, player1's movement is very choppyand lags behind player1's actual position.

从玩家 2 的角度来看，玩家 1 的动作非常不稳定，落后于玩家 1 的实际位置。

Interpolation & Extrapolation

内插值和外插值

Purpose: Smooth movementof player's characters on screen

目的：玩家角色在屏幕上流畅移动

Interpolation

内插值

Calculate the state between old butknown states

计算旧但已知状态之间的状态

Extrapolation

外插值

Predict where entity is going from old states

预测实体从旧状态走向何方

Smooth States by Interpolations

插值平滑状态

Position and Orientationcan be interpolated between two recently received data

位置和方向可以在最近接收到的两个数据之间进行插值

Buffer States and Deferred Render

缓冲区状态和延迟渲染

Data packet will not be rendered immediately when received

数据包在收到后不会立即呈现
Put into memory and wait for a new data packet

放入内存并等待新的数据包
After waiting for a time offset, start to render first received data packet

等待时间偏移后，开始渲染第一个接收到的数据包
Create an artificial delay of interpolation offset

创建插值偏移的人工延迟

Character Movement Replication by Interpolation

通过插值同步角色动作

Result after interpolation was implemented

插值后的结果

Interpolation Challenges of Vehicle Movement Replication

车辆运动同步的插值挑战

Estimate Current State by Extrapolation

通过外推法估计当前状态

Use paststates to estimate currentstate to compensate net lag

使用过去状态来估计当前状态，以补偿净滞后

Dead Reckoning

死亡清算

Estimate future state based on states that have been received

根据已收到的状态估计未来状态

Projective Velocity Blending

投影速度混合

At $t_0$, the replicated character is at $p_0$ with velocity $v_0$ and acceleration $a_0$, and receive the
synced states with position $p'_0$, velocity $v'_0$, acceleration $a'_0$

在时间 $t_0$，同步的角色位于 $p_0$，其速度为 $v_0$，加速度为 $a_0$，并接收到同步的状态，其位置为 $p'_0$，速度为 $v'_0$，加速度为 $a'_0$。
We can predict position $p'_t$ after a time duration $t$ based the synced states

我们可以根据同步的状态预测在时间持续 $t$ 后的位置 $p'_t$。

$p'_t=p'_0+v'_0t+\frac{1}{2}a'_0t^2$

Our goal is to reach $p'_t|_{t=t_B}$ smoothly after $a$ fixed blending time duration: $t_B-t_0$

我们的目标是在固定的混合时间持续 $t_B - t_0$ 后平滑地到达 $p'_t |_{t=t_B}$。

At any time $t$, we can get the blending velocity $v_t$

在任何时刻 $t$，我们可以得到混合速度 $v_t$

$\lambda=\frac{t-t_0}{t_B-t_0}$ $v_t=v_0+\lambda(v'_0-v_0)$

And projecting the position $p_t$ from $p_0$

并从 $p_0$ 投影位置 $p_t$

$p_t=p_0+v_tt+\frac{1}{2}a'_0t^2$

Then get the dead reckoned position $p_d$ by combining $p_t$ and $p'_t$

然后通过结合 $p_t$ 和 $p'_t$ 得到估算位置 $p_d$

$p_d=p_t+\lambda(p'_t-p_t)$

Collision Issues

碰撞问题

Dead reconking Collision trajection looks weird

估算碰撞轨迹看起来很奇怪

Phase 1: Collision starts

第一阶段：碰撞开始

Phase 2: The replica keeps going, since the extrapolation is basedon the last snaphsot

第二阶段：同步继续进行，因为外推是基于最后一次快照

Phase 3: Finally we receive a snapshot to stop the replica, but replica gives master’s rigidbody a huge velocity to pushing master away

第三阶段：最后我们收到一个快照来停止同步，但同步会给主体的刚体以巨大的速度将主体推开

Physics Simualtion Blending During Collision

碰撞过程中的物理模拟混合

Tunable between two states

可在两个状态之间调整

State calculated by the client physics simulation

客户端物理模拟计算的状态
State that tries to reach the dead reckoned positions

试图到达指定位置

Tuned blending factors from Watch Dogs 2, Ubisoft Toronto. Bikes recover faster than cars

调整了育碧多伦多《看门狗 2》的混合因子。自行车的恢复速度比汽车快。

Usage Scenario of Interpolation

插值使用场景

Scenario for Using Interpolation

使用插值的场景

Characters' movement are very non-deterministic with high acceleration

角色在高加速度下的运动非常不确定
Gameplay suffers from the “wrap”when extrapolation errors occur

当出现外推错误时，游戏玩法会受到“包装”的影响

Typical examples

典型示例

FPS
MOBA

Usage Scenario of Extrapolation

外推法使用场景

Scenario for Using Extrapolation

使用外推法的场景

Player movement uses a realistic physical model

玩家动作使用逼真的物理模型
Gameplay suffers from latency due to network transmission

游戏因网络传输而延迟

Typical examples

典型示例

Racing game. Vehicle systems (Tanks, Ships, etc.)

赛车游戏。载具系统（坦克、船舶等）

Blend Scenario of Interpolation and Extrapolation

插值和外推的混合场景

Sometimes we need to apply both interpolation and extrapolation for the game to work properly

有时我们需要同时应用插值和外推法才能使游戏正常工作

Apply Extrapolation on vehicles

对车辆应用外推法
Apply Interpolation for characters

对角色应用插值
Do extrapolation if not enough data received

如果没有收到足够的数据，则进行外推

Hit Registration

点击注册

How to Make a Headshot in Online Game

如何在网络游戏中进行头部射击

Net messages to travel from client to server, and interpolation causes you to see the enemy way lag behind

网络消息从客户端传输到服务器，插值会导致你看到敌人的方式落后

Where is the Enemy?

敌人在哪里？

Due to latency, interpolation offsetand time delay, you'll see other players slightly behind their current server positions. Where should I shot?

由于延迟、插值偏移和时间延迟，你会看到其他玩家稍微落后于他们当前的服务器位置。我应该在哪里开枪？

Where Should I Shot?

我应该在哪里开枪？

Hit Registration

点击注册

Hit registrationis making a consensusof all players that whether you've actually hit your enemy

命中注册是让所有玩家就你是否真的击中了敌人达成共识。

Detecting hit event on client-side with replicated character positions

通过同步角色位置在客户端检测命中事件
Send hit events to the server

将命中事件发送到服务器
Server run simple verification

服务器运行简单验证

A Comparison of HitscanWeapons versus Projectile Weapons

Hitscan 武器与投射武器的比较

Unlike hitscanweapons, projectile weapons can also simulate the effect of gravity

与 hitscanweapons 不同，投射式武器还可以模拟重力的影响

《战地》中的风景是由几个 hitbox 构建的，因此破坏可以带走墙壁、地板等。

Hitscan 发射射弹后立即击中目标
Projectile 发射射弹后需要一定时间才可击中目标

A Very Simple Server Verification of Hit Event

一种非常简单的服务器命中事件验证

Client send hit event with complete ray information to server

客户端向服务器发送包含完整射线信息的命中事件
- StartPoint, HitPointand HitObjectof the Raycast
  
  光线投射的起点、终点和终点
Validate StartPointwhether is really close enough to shooter

验证起点是否真的足够接近射击者
Validate the HitPointwhether is really belong to HitObject

验证 HitPoint 是否真的属于 HitObject
Ensure nothing is blocking along the path by casting a ray from the StartPointand HitPoint

通过从起点和终点投射光线，确保路径上没有任何遮挡

In real game, the server verification is VERY TRICKY AND COMPLICATED

在真实游戏中，服务器验证非常棘手和复杂

Server Verification Has to Guess

服务器验证必须猜测

Problem of Client-Side Hit Detection

客户端命中检测问题

Efficient and Precise

高效精准

Very efficient for hit detection without huge server workload

非常有效的命中检测，无需巨大的服务器工作负载
Best shooting experience with pixel precision

像素精度最佳的拍摄体验

Unsafe for cheating

易于玩家作弊

Fake hit event message

假热门事件消息
Lag switches

滞后开关
Infinite ammo

无限弹药
…

Detecting Hit on Server-Side?

检测服务器端的命中？

Client doesn’t know the target current location on server

客户端不知道服务器上的目标当前位置

Lag Compensation

滞后补偿

Server-side state rewinding to compensate network lags when player’s commands are executed

服务器端状态倒带，以补偿执行玩家命令时的网络延迟

Get information from clients

从客户那里获取信息
Rewind game state in cached state snapshots that matches the client's action time

在与客户端操作时间匹配的缓存状态快照中回放游戏状态
Run client operation in rewind game state

在倒带游戏状态下运行客户端操作

Compensate all Possible Lags

补偿所有可能的延迟

RewindTime = Current Server Time - Packet Latency - Client View Interpolation Offset

RewindTime = 当前服务器时间 - 数据包延迟 - 客户端视图插值偏移

Actor: Enemy's client state

角色：敌人的客户端状态

Red collision box: Enemy in the player's view

红色碰撞框：玩家视野中的敌人

Blue collision box: Rewinded server state

蓝色冲突框：倒带服务器状态

Cover Problems –Running into Cover

掩体问题——遇到掩体

被击中者躲到掩体后因为网络延迟仍被击中，Shooter' s advantage.

Cover Problems –Coming out from Cover

被攻击者从掩体出来因网络延迟未被击中，Peeker’s advantage.

Startup Frames to Ease Latency Feeling

启动框架缓解延迟感

A fixed animation before attack or move can also eliminate the effect of lag from network transmission

攻击或移动前的固定动画（技能前摇）也可以消除网络传输延迟的影响
Players will keep their attention on animations and ignore the state delay

玩家将注意力集中在动画上，忽略状态延迟

Local Forecast VFX Impacts

本地预测 VFX 影响

Clients can perform local hit tests in order to give the player some instant feedback, such as displaying a blood splatter visual effect

客户端可以执行本地命中测试，以便为玩家提供一些即时反馈，例如显示血溅的视觉效果
However, any permanent effects of the hits, such as reducing the hit points of a player, are only applied after receiving confirmation from the server

但是，只有在收到服务器的确认后，才会应用命中的任何永久效果，例如减少玩家的命中点

MMOG Network Architecture

MMOG 网络架构

What is MMOG?

什么是 MMOG？

MMOG: Massively Multiplayer Online Game, or more commonly MMO

MMOG：大型多人在线游戏，或更常见的 MMO

MMOs with a large numbers of players, often hundreds or thousands, on the same server can enable players to cooperate and compete with each other on a large scale, and include a variety of gameplay types (MMORPG, MMORTS, MMOFPS, etc.)

在同一台服务器上拥有大量玩家（通常是数百或数千人）的 MMO 可以使玩家能够大规模地相互合作和竞争，并包括各种游戏类型（MMORPG、MMORTS、MMOFPS 等）

Diversities of Modern MMO

现代 MMO 的多样性

Game Sub-Systems

游戏子系统

MMOs have a variety of gameplay and are supported by many sub-systems

MMO 具有多种游戏玩法，并得到许多子系统的支持

User management

用户管理
Matchmaking

配对
Trading system

交易系统
Social system

社会制度
Data storage

数据存储
...

MMO Architecture

MMO 架构

Services of Link Layer

链路层服务

登录服务器

Verification of client connection

客户端连接验证

Gateway

网关

Very important layer to separate inside/outside networks

分隔内部/外部网络的非常重要的层

Lobby

大厅

Players can gather in the lobby, see and interact with other players

玩家可以聚集在大厅，观看并与其他玩家互动
When the number of players continues to increase, it is a challenge to the performance of the server and the client

当玩家数量持续增加时，服务器和客户端的性能都会受到挑战

Character Server

角色服务器

All player data is managed in one system. Such as account info, character info, backpack info, mail info, etc.

所有玩家数据都在一个系统中管理。如账户信息、人物信息、背包信息、邮件信息等。

Trading System

交易系统

Buying and selling items on the marketplace

在市场上买卖商品
Sending items or coins to other players through the in-game Mail

通过游戏内邮件向其他玩家发送物品或硬币
Game designers need to keep an eye on market prices to prevent imbalances

游戏设计师需要关注市场价格，以防止不平衡
For a persistent world to maintain a stable economy, a balance must be struck between currency sources and sinks

为了使世界保持稳定的经济，必须在货币来源和汇之间取得平衡
Players can use real-world money to buy a specific in-game item

玩家可以用现实世界的钱购买特定的游戏内物品

社会制度

Player-to-player interplay and communication

玩家之间的互动和沟通
Foster stronger social cohesion in-game

在游戏中培养更强的社会凝聚力

Matchmaking

配对

You have to consider attributes like skills, level, latency, wait time...

你必须考虑技能、级别、延迟、等待时间等属性…
Ingeneral, makingagood matchmaking service is core for a game design

一般来说，提供良好的配对服务是游戏设计的核心
Running this on a global scale for your player population presents a whole different set of challenges

在全球范围内为您的玩家群体运行此功能会带来一系列完全不同的挑战

Data Storage

数据存储

The game data is very complex and diverse

游戏数据非常复杂和多样化

Player data (guilds, dungeons, warehouse, etc.)

玩家数据（公会、地下城、仓库等）
Monitoring data

监测数据
Mining data

最小化数据
...

Data needs to be securely persisted and efficiently organized for retrieved and analysis etc.

数据需要安全地持久化，并有效地组织起来进行检索和分析等。

Relational Data Storage

关系数据存储：MySQL

Requires Structure to be Predetermined

需要预先确定结构
Flexible Queries

灵活的查询
Always Consistent

始终保持一致

Game Development Examples

游戏开发示例

Player Data

玩家数据
GameData

游戏数据
Inventory

库存
ItemShops/Trading

商品商店/贸易

Non-Relational Data Storage

非关系数据存储：MongoDB

Structure Can Change For Each Entry

每个条目的结构都可能发生变化
Queries Have Higher Specificity

查询具有更高的特异性
May Not Always Be Consistent

可能并不总是一致的

Game Development Examples

游戏开发示例

Player/Item Stats/Profile Game Data

玩家/物品统计/个人资料游戏数据
Enchantments and Upgrades

魔法和升级
Game States

游戏状态
Quest Data

任务数据

In-Memory Data Storage

内存数据存储：redis

Extremely Fast (Memory versus Hard Disk)

极快（内存与硬盘）
Key-Value

关键价值
Fast Sorted/Ranged Searches

快速排序/范围搜索
Persistence among servers

服务器间的持久性

Game Development Examples

游戏开发示例

Matchmaking

配对
Leaderboards

排行榜
Session Management

会话管理
Boost Performance For Other Databases

提高其他数据库的性能

Player Number Growth

玩家数量增长带来的服务器负担

Distributed System

分布式系统

A distributed system is a computing environment in which various components are spread across multiple computers (or other computing devices) on a network

分布式系统是一种计算环境，其中各种组件分布在网络上的多台计算机（或其他计算设备）上

Challenges with Distributed systems

分布式系统的挑战

Data access mutual exclusion

数据访问互斥
Idempotence

意识
Failure and partial failure

故障和部分故障
Unreliable network

不可靠的网络
Distributed bugs spread epidemically

分布式漏洞以流行病的方式传播
Consistency and consensus

一致性和共识
Distributed transaction

分布式事务

Load Balancing

负载平衡

Refers to the process of distributing a set of tasks over a set of resources (computing units), with the aim of making their overall processing more efficient

指将一组任务分配到一组资源（计算单元）上的过程，目的是提高其整体处理效率

Optimize the response time

优化响应时间
Avoid unevenly overloading some compute nodes while other compute nodes are left idle

避免在其他计算节点处于空闲状态时，不均匀地使某些计算节点过载
All players are evenly divided on multiple servers

所有玩家平均分布在多个服务器上

Consistent Hashing

一致散列

It was designed to avoid the problem of having to reassign every player when a server is added or removed throughout the cluster

它旨在避免在整个集群中添加或删除服务器时必须重新分配每个玩家的问题

Virtual Server Node in Consistent Hashing

一致散列中的虚拟服务器节点

Servers Management

The number of services increases

服务数量增加
Difficult to manage

难以管理
Lacks the flexibility to change the IP or port at a later point in time

缺乏在以后更改 IP 或端口的灵活性

Service Discovery - Registry

服务发现-注册表

Registers itself with the service registry when it enters the system

当服务进入系统时，在服务注册表中注册
An example of Register value

寄存器值示例
- server type/server_name@server_ip:port

Service Discovery - Query and Watch

服务发现-查询和监视

Request service discovery service to query all values through service type and watch it

请求服务发现服务通过服务类型查询所有值并查看

Service Discovery - Health Check

服务发现-健康检查

Notice Gateway Server B Failure when Server Instance B Heartbeat timeout

服务器实例 B 心跳超时时通知网关服务器 B 故障

Bandwidth Optimization

优化带宽

Why Bandwidth Matters

为什么带宽很重要

Usage-based billing: e.g. mobile, cloud service

基于使用量的计费：例如移动、云服务
Latency increased by bandwidth: packet splitting/drop

延迟随带宽增加：数据包拆分/丢弃
Connection drops due to message overflow

消息溢出导致连接中断

Calculate Bandwidth

计算带宽

Affecting factors

影响因素

n = player numbers 玩家数量
f = update frequency 更新频率
s = size of game state 游戏状态大小

Data transfer per second

每秒数据传输

Server 服务器：$O(n\cdot s\cdot f)$
Client (downstream) 客户端（下游）：$O(s\cdot f)$
Client (upstream) 客户端（上游）：$O(f)$

Data Compression

数据压缩

There are a lot of floating point numbers in the game synchronization data, such as position, rotation, speed, etc.

游戏同步数据中有很多浮点数，如位置、旋转、速度等。
Choosing the right floating-point precision can save a lot of bandwidth

选择正确的浮点精度可以节省大量带宽。
- e.g When representing human running speed, only half precision is required
  
  例如，在表示人类跑步速度时，只需要一半的精度

When representing player position, the player will only move within a certain range due to player speed limitations

代表玩家位置时，由于玩家速度限制，玩家只能在一定范围内移动
We can divide the map into different small pieces and use the relative position to represent the player's position, which can reduce the precision of the floating point number of the synchronization position

我们可以将地图分成不同的小块，并使用相对位置来表示玩家的位置，这会降低同步位置浮点数的精度

Object Relevance

对象相关性

The player will be informed of state updates

玩家将收到状态更新的通知
Usually, the ones player can see & interact

通常，玩家可以看到并互动
Easiest implementation: all objects relevant to all clients (for small player num).

最简单的实现：与所有客户端相关的所有对象（对于小玩家 num）。$O(n^2)$
Limiting factor for max concurrent players

最大并发玩家的限制因素

Relevance - Static Zones

Relevance - Area of Interest (AOI)

AOI -Direct Range-Query

AOI - 直接范围查询

$\sqrt{(x_{player}-x_i)^2-(y_{player}-y_i)^2}\le r_{doi}$
Simple to implement 易于实施
Time complexity 时间复杂度: $O(n^2)$
Not suitable for MMOG, e.g. 1000 players in one zone, 20 ticks/s

不适合 MMOG，例如一个区域有 1000 名玩家，20 滴答/秒

1000 x 1000 x 20 = 20,000,000 distance computations per second

1000 x 1000 x 20 = 每秒 20000000 次距离计算

AOI - Spatial-Grid

AOI - 空间网格

Mapping Entities

映射实体

Map entity $(x, y)$ → grid $N$

映射实体$(x, y)$ → 网格 $N$
Relevant entities in the grids around current player's grid

当前玩家网格周围网格中的相关实体
Player's AOI list can be cached

玩家的 AOI 列表可以缓存

Events

事件

Enter

进入

Add entities from observation (observed) list

从观察（observation）列表中添加实体

Leave

离开

Remove entities from observation (observed) list

从观察（已观察）列表中删除实体

Pros and Cons

优点和缺点

Pros

Fast query time $O(1)$

快速查询时间 $O(1)$

Cons

Small grid: high memory cost

小网格：内存成本高
Large grid: high CPU cost

大网格：CPU 成本高
Object with varying AOI radius?

具有不同 AOI 半径的物体？

AOI - Orthogonal Linked-list

AOI - 正交链表

Game entities in double linked-list

双链表中的游戏实体
- xlist, ylist
- ascending order
  
  升序
Less Objects to traverse

需要遍历的对象更少

Traverse entities

遍历实体

Within aoi radius

aoi 半径内
Left/right direction

左 / 右方向
For both x/y lists

对于 x/y 两个列表

Better Approach - Range Trigger

更好的方法-范围触发

Entity move → trigger move

实体移动 → 触发移动
Compare with trigger

与触发器比较
Event driven

事件驱动

Pros

Memory efficient

内存效率高
Varying AOI radius

可变 AOI 半径

Cons

New object insertion cost $O(n)$

新对象插入成本 $O(n)$
Not Suitable when entities move large distance frequently

当实体频繁移动远距离时不适用

AOI - Potentially Visible Set (PVS)

AOI - 潜在可见集（PVS）

Set of potentially visible areas

一组潜在可见区域
Can be calculated offline

可以离线计算
Determine relevant objects from PVS

从 PVS 中确定相关对象
e.g. Racing game: fast-moving car

例如赛车游戏：快速行驶的汽车

Varying Update Frequency by Player Position

根据玩家位置改变更新频率

Distance-based update frequency

基于距离的更新频率
Only closer objects are interactable

只有较近的对象是可交互的
Distance ↑ → f → bandwidth ↓

距离 ↑ → f → 带宽 ↓

Cheating Kill Online Games

作弊杀死在线游戏

How likely, if at all, would you be to stop playing a multiplayer game online if you thought other players were cheating to gain an unfair advantage?

如果你认为其他玩家作弊以获得不公平的优势，你有多大可能停止玩在线多人游戏？

	Very likely	Fairly likely	Not very likely	Nor linkly at all	Don't konw
Global	29%	48%	15%	4%	5%
China	25%	56%	16%	2%	1%
Germany	30%	36%	17%	8%	8%
Japan	26%	49%	16%	2%	7%
South Korea	27%	59%	11%	1%	1%
UK	33%	39%	12%	6%	10%
US	37%	32%	15%	7%	8%

77% of players will likely stop playing online games when other players are cheating, according to the survey of Irdeto.

爱迪德的调查显示，当其他玩家作弊时，77% 的玩家可能会停止玩网络游戏。

Millions Ways of Cheating

数百万种作弊方式

Game code modifications

游戏代码修改

Modify or read memory data

修改或读取内存数据
Crack client

破解客户端

System software invoke

系统软件调用

D3D Render Hook

D3D 渲染挂钩
Simulate mouse and keyboard operations

模拟鼠标和键盘操作
...

Net Packet interception

网络数据包拦截

Send fake packets

发送虚假数据包
Modify packet data

修改数据包

Obfuscating Memory

混淆记忆

A cheater might be able to get the location of the player coordinates in the memory and move the character ignoring the game rules, such as passing the wall

作弊者可能能够获得玩家在内存中的坐标位置，并忽略游戏规则移动角色，例如通过墙壁
Furthermore, the cheater can utilize the location of these values to map out even larger data structures in the memory, such as the player object itself

此外，作弊者可以利用这些值的位置在内存中绘制出更大的数据结构，例如玩家对象本身

Executable Packers

可执行打包器

Game core logic can be restored by reverse engineering

游戏核心逻辑可以通过逆向工程恢复
Players can crack the game by analyzing the code, finding game loopholes, making plug-ins, etc..

玩家可以通过分析代码、发现游戏漏洞、制作插件等来破解游戏…

The packager obfuscates the source program and adds decompression code

打包器混淆源程序并添加解压缩代码
The decompression code will execute first, and the source program is decrypted in memory

解压缩代码将首先执行，源程序在内存中解密

Verifying Local Files by Hashing

通过哈希验证本地文件

Ensure that the game files have not been modified

确保游戏文件未被修改
For example, the cheater could modify the wall textures to be transparent so all enemies could be seen through the walls

例如，作弊者可以将墙壁纹理修改为透明，这样所有敌人都可以透过墙壁看到
The cheater could also adjust the lightning to make it easier to see enemies

作弊者还可以调整闪电，使其更容易看到敌人

Packet Interception and Manipulation

数据包拦截和操纵

When the data is not encrypted or hacked, the player can build game logic based on packet data even without starting the game

当数据未被加密或黑客攻击时，即使不启动游戏，玩家也可以基于数据包构建游戏逻辑
Such cheat programs often become money-making tools, which seriously reduce game's the overall profit

这种作弊程序往往成为赚钱的工具，严重降低了游戏的整体利润

Encrypt the Network Traffic

加密网络流量

Two kinds of algorithms

两种算法

Symmetric-key algorithm

对称密钥算法
- Obfuscate and restore data according to the same key
  
  根据同一密钥混淆和还原数据
- Fast and efficient
  
  快速高效

Asymmetric encryption

非对称加密

Encryption and decryption use different keys

加密和解密使用不同的密钥
Slow, only used for encrypting critical data

速度慢，仅用于加密关键数据

Distribute symmetric key securely using asymmetric encryption

使用非对称加密安全地分发对称密钥
Transfer data using symmetric encryption key

使用对称加密密钥传输数据

System Software Invoke

系统软件调用

Modify the DirectX kernel and change the execution flow of the rendering function

修改 DirectX 内核并更改渲染函数的执行流程
Can force the rendering engine to modify the occlusion relationship

可以强制渲染引擎修改遮挡关系
See the movement of the enemy behind the wall

看墙后敌人的动向

Valve Anti-Cheat and Easy Anti-Cheat

阀门防作弊和简易防作弊

Detects malicious behavior caused by any file conflicts while interacting with the game

检测与游戏交互时由任何文件冲突引起的恶意行为
Stops the player from playing the game at all

完全阻止玩家玩游戏
Prevents any illegal modifications and configuration changes that enable the use of exploits in a game

防止任何非法修改和配置更改，从而在游戏中使用漏洞

AI Cheat

AI 作弊

All platforms

所有平台
No code modification required

无需修改代码
Independent from the game

独立于游戏
Game screen

游戏画面
Target detection

目标检测
Move cursor

移动光标
Fire

开火

Rich AI Middlewares

丰富的 AI 中间件

Real-Time Object Detection. YOLO V5, V7...

实时目标检测。YOLO V5，V7…
Skeleton based Action recognition

基于骨架的动作识别

Counter-Strike: Overwatch

反恐精英：守望先锋

The system is based on other players reviewing footage from players that are suspected of cheating

该系统基于其他玩家查看涉嫌作弊的玩家的录像
Many reviewers are looking at the same cases and the majority decide whether the suspect was cheating or not

许多审查人员正在审查同样的案件，大多数人决定嫌疑人是否作弊

Passing judgement after reviewing evidence in Counter Strike: Global Offensive’s Overwatch system

在《反恐精英：全球攻势的守望先锋》系统中审查证据后作出判断

Statistic-based System

基于统计的系统

Collect the user's game information, such as victory and critical hit rate

收集用户的游戏信息，如胜利和暴击率
Compare your own historical data and some thresholds rules or from other player's reports to mark players

比较您自己的历史数据和一些阈值规则或其他玩家的报告，以标记玩家
Check manually to confirm whether they cheat

手动检查以确认他们是否作弊

Detecting Known Cheat Program

检测已知作弊程序

A proper anti-cheat program should have a way to scan the user's computer for known cheating programs based on various signatures

一个合适的反作弊程序应该有一种方法，可以根据各种签名扫描用户的计算机，寻找已知的作弊程序
The simplest method can simply entail comparing hashes or process names

最简单的方法可能只是比较哈希或进程名称

Build a Scalable World

构建一个可扩展的世界

Scalable Game Servers

可扩展游戏服务器

Zoning

分区

Distribute large player numbers in a large world

在广阔的世界中分配大量玩家
Distribution might be uneven

分布可能不均匀

Instancing

实例化

Run a large number of game areas independently in parallel

并行独立运行大量游戏区域
Reduce congestion/competition

减少拥堵 / 竞争

Replication

同步

Allows high user density

允许高用户密度
E.g. high density PVP games

例如高密度 PVP 游戏

Zoning - Seamless Zones

分区-无缝分区

Players are reasonably distributed in a large world

玩家在广阔的世界中分布合理
The client only connects to one responsible server

客户端只连接到一个负责的服务器
Cross border: auto transfer client to another server

跨境：自动将客户端转移到另一台服务器

Zone Border

区域边界

Smooth experience:

流畅体验：

Border width >= max AOI radius

边框宽度 >= 最大 AOI 半径

But how to make them interact?

但是如何让它们互动呢？

Zone Border - Entities

区域边界-实体

Active Entity

活跃实体

Resides in connected zoned server (authority)

驻留在已连接的分区服务器（权限）中
Has a ghost agent in other zones

在其他区域有幽灵特工
Can see ghost entities in another zone

可以在另一个区域中看到幽灵实体

Ghost Entity

幽灵实体

Also called shadow entity

也称为阴影实体
Is an agent entity owned by another zone

代理实体是否由另一个区域拥有
Receive updates from original entity

接收来自原始实体的更新

Cross Border: A -> B

跨境：A->B

Before move

移动前
- An active entity in zone A
  
  A 区中的活动实体
Near boundary (A)

近边界（A）
- Active in A; Ghost in B
  
  活跃在 A；鬼魂在 B
At boundary

边界处
- The entity has been transferred to zone B
  
  实体已转移到 B 区
Near boundary (B)

边界附近（B）
- Active in B; Ghost in A
  
  活跃于 B；鬼魂在 A
Beyond boundary (B)

边界之外（B）
- Removedfrom zone A
  
  从 A 区移除

Replication

同步

Cooperatively process same world zone

合作处理同一世界区
Entity updates are distributed among servers

实体更新分布在服务器之间
Each server creates its own active entities

每个服务器都创建自己的活动实体
Updates to active entities will be auto replicated to all remaining servers (as Ghost)

对活动实体的更新将自动复制到所有剩余的服务器（作为 Ghost）

Scalable Game Servers -Combination

References

Replicate Character Movement

Replicating Chaos: Vehicle Replication in 'Watch Dogs 2', Matt Delbosc, Ubisoft Toronto, GDC 2017: https://www.gdcvault.com/play/1024597/Replicating-Chaos-Vehicle-Replication-in
"Believable dead reckoning for networked games.", Murphy, Curtiss, and E. Lengyel, Game Engine Gems 2 (2011) : 307-328. https://www.researchgate.net/publication/293809946_Believable_Dead_Reckoning_for_Networked_Games
Client-side Interpolation:https://docs-multiplayer.unity3d.com/netcode/0.1.0/learn/clientside_interpolation/index.html
Source Multiplayer Networking:
https://developer.valvesoftware.com/wiki/Source_Multiplayer_Networking

Lag Mitigation

Latency Compensating Methods in Client/Server In-game Protocol Design and Optimization, Yahn W. Bernier: https://www.gamedevs.org/uploads/latency-compensation-in-client-server-protocols.pdf
Valorant's netcode:https://technology.riotgames.com/news/peeking-valorants-netcode
Source Multiplayer Networking, valve developer community: https://developer.valvesoftware.com/wiki/Source_Multiplayer_Networking
Implementation and Evaluation of Hit Registration in Networked First Person Shooters, Jonathan Lundgren:https://liu.diva-portal.org/smash/get/diva2:1605200/FULLTEXT01.pdf
How It Works: Lag compensation and Interp in CS:GO:https://www.youtube.com/watch?v=6EwaW2iz4iA&ab_channel=DevinDTV

MMOG Network Architecture

Intro to Databases in Games: How to Use Them in Games and Game Development -AWS Online Tech Talks: https://www.youtube.com/watch?v=7HppNxu_hdA
Massively multiplayer online game, WIKI: https://en.wikipedia.org/wiki/Massively_multiplayer_online_game
Consistent Hashing and Random Trees:
Distributed Caching Protocols for Relieving Hot Spots on the World Wide Web: https://dl.acm.org/doi/pdf/10.1145/258533.258660
Consistent hashing:https://en.wikipedia.org/wiki/Consistent_hashing
Fowler–Noll–Vo hash function: https://en.wikipedia.org/wiki/Fowler%E2%80%93Noll%E2%80%93Vo_hash_function
Performance Testing Methodology: http://hosteddocs.ittoolbox.com/questnolg22106java.pdf
Glinka, F.et al. (2008) 'High-Level Development of Multiserver Online Games', International Journal of Computer Games Technology,2008(327387). Available at: https://downloads.hindawi.com/journals/ijcgt/2008/327387.pdf
Alibaba Cloud MMO Gaming Solution Architecture: https://www.alibabacloud.com/blog/593877

Bandwidth Optimization

Replication in network games: Bandwidth (Part 4): https://0fps.net/2014/03/09/replication-in-network-games-bandwidth-part-4/
Introducing Time Dilation: https://www.eveonline.com/news/view/introducing-time-dilation-tidi
Glazer, J. and Madhav, S. (2016) Multiplayer Game Programming: Architecting Networked Games. New York: Addison-Wesley
Managing and Mining Multiplayer Online Games for the Summer Semester 2017: https://www.dbs.ifi.lmu.de/Lehre/mmmo/sose17/slides/MMMO-2-Core_1.pdf
Handling large amounts of players: https://mmo-blueprint.com/handling-large-amounts-of-players/?i=1
Zinx应用-MMO游戏案例: https://github.com/aceld/zinx
KBEngine: https://github.com/kbengine/kbengine

Anti-Cheat

Cheating in online games, WIKI: https://en.wikipedia.org/wiki/Cheating_in_online_games
The Day FPS Games Died, 2021.07: https://www.youtube.com/watch?v=revk5r5vqxA
警惕 AI 外挂, LYi, 2021.08: https://www.bilibili.com/video/BV1Lq4y1M7E2
YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors, Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao, 2022.07: https://arxiv.org/abs/2207.02696
MarI/O -Machine Learning for Video Games, SethBling: https://www.youtube.com/watch?v=qv6UVOQ0F44
Comparative Study of Anti-cheat Methods in Video Games, Samuli Lehtonen, 2020.3:https://helda.helsinki.fi/bitstream/handle/10138/313587/Anti_cheat_for_video_games_final_07_03_2020.pdf
游戏外挂攻防艺术, 徐胜, 2013
Python Plays Grand Theft Auto 5 -Reboot, Harrison: https://github.com/Sentdex/pygta5

Citation

Maze War (1978) -First Person Shooter History -Ep01: https://www.youtube.com/watch?v=ZjY5s_05Qlw
How It Works: Lag compensation and Interp in CS:GO:https://www.youtube.com/watch?v=6EwaW2iz4iA&ab_channel=DevinDTV
World of Warcraft (2021) -Gameplay (PC UHD) [4K60FPS]: https://www.youtube.com/watch?v=d26Kl98_qo8
Guild Wars 2’s greatest STRENGTH is also its WEAKNESS: https://www.youtube.com/watch?v=jGcQDhSPLHM
FINAL FANTASY XIV: https://jp.finalfantasyxiv.com
WHY PLAY EVE ONLINE in 2022? -And why this is one of my favorite games: https://www.youtube.com/watch?v=j2t_XPDOIFc
World of Warcraft: Cataclysm Login Screen: https://www.youtube.com/watch?v=isCsAEwpIOY
10 Best MMOs To Play In 2022: https://www.youtube.com/watch?v=W0rvIv6diVU
Planetside 2: https://www.planetside2.com/home
Planetside 2: Things Just Got Harder: https://www.youtube.com/watch?v=HTN-yc5x0pA
15 Years of WoW vs 1 Year of FFXIV: https://www.youtube.com/watch?v=5T-mcLYBnKc
MUD1 (1978) -First MUD game:https://www.youtube.com/watch?v=9Gep3LwLKWk