Alibaba final interview: How to use UDP to implement TCP?

[[355616]]

This article comes from a real interview experience of mine.

background

This question was asked by the technical director during my interview. I didn’t answer it very well at the time, so I summarized it. Later, I asked the director, and he said it was an interview question from Alibaba.

In fact, the interviewer mainly wanted me to explain the difference in principles between UDP and TCP, and how to add some functions to UDP to implement TCP.

It's easy to tell one or two differences between TCP and UDP, but how can you say it in a way that your girlfriend can understand?

Girlfriend: I don’t want to listen to what’s in the textbook! I don’t understand.

Below I will answer the above questions in plain language.

Characteristics of UDP

UDP reminds me of when I was a fresh graduate and started working.

• Simple communication

Just do the tasks assigned by your leader and complete them.

UDP is also true. I believe that the network world will always be beautiful. The packets I send are easy to deliver and the receiver is also easy to assemble. The data structure is also simple, and does not require a large amount of data structure, processing logic, or packet header fields.

• Trust others

I will not argue with the tester about the bugs reported by her. I will always believe that the tester is right. Whatever the tester says is what I will fix.

UDP is the same. It does not establish a connection. It has a port number. Anyone can listen to this port number and send data to it. It can also send data to anyone from this port number. Anyway, I just send it.

• No bargaining

The product manager said yesterday that the phone case needs to change color according to the mood, and the tester said that this bug needs to fix two related bugs together. So let's do what they say!

UDP is the same, it doesn't know how to persist and give in. That is, it performs congestion control according to the network conditions. No matter how serious the network packet loss is, I will still send it.

UDP usage scenarios

In view of the situation of a new graduate like me at that time, my leader arranged three working environments for me to choose from.

• The internal system has simple tasks and a single module. There is no need to consider the associated impact of the code, and it does not matter even if it fails.

UDP also requires fewer resources, an intranet with relatively good network conditions, or applications that are not sensitive to packet loss.

• There is a strong team support, all of whom are mid-level and senior developers and testers. The team members have worked together for many years and trust each other. If you have any problems, just shout!

UDP also does not require one-to-one communication to establish a connection and can be used for broadcasting applications.

• A new project requires passion. For rookies who have just graduated, they are very self-motivated and will not play tricks, hide in the toilet to play with their phones, or take paid shits. Even if the project is not busy, I will seize the time to work. Even if the project is busy, I will still work!

UDP is also like this, it sends packets quickly, and is mainly used in situations where fast processing speed, low latency, and a small amount of packet loss can be tolerated. Even if the network conditions are not good, it just sends packets~

For the above three scenarios, UDP is often used in real-time competitive games, IoT, and mobile communications.

What are the characteristics of TCP?

• Connection-Oriented

TCP and UDP are two important protocols in the transport layer. Most interviews will ask about the difference between the two. Most people will answer two sentences, such as TCP is connection-oriented and UDP is connectionless.

So what is connection-oriented?

The TCP three-way handshake is what we often talk about and recite. After these three handshakes are successful, the connection is successfully established.

So what is orientation?

We often hear about object-oriented programming, aspect-oriented programming, and service-oriented programming. So what exactly is object-oriented programming?

In my opinion, orientation means following certain protocols, specifications, data structures, etc. to do a series of things.

For example, connection-oriented technology is to maintain the connection between the client and the server, and to establish a certain data structure to maintain the status of the interaction between the two parties. Such data is used to ensure the so-called connection-oriented characteristics.

Knowing that TCP uses three-way handshake to establish a connection, can we let UDP also send three packets to simulate TCP to establish a connection? Of course, we can, but if it is just to establish, not to be connection-oriented, it is actually not very meaningful.

So what does TCP connection-oriented do?

TCP provides reliable delivery. Data transmitted through a TCP connection can be error-free, non-lost, non-duplicate, and arrive in order. However, UDP inherits the characteristics of IP packets and does not guarantee non-loss or in-order delivery.

• Byte stream oriented

TCP is oriented to byte streams. The so-called byte stream is a stream that has no beginning or end. TCP maintains the stream state itself.

UDP is based on IP datagrams, which are sent and received one by one.

• Congestion Control

TCP has congestion control. If a packet is discarded or the network environment is bad, it will control its own behavior according to the network conditions to see whether to send faster or slower.

UDP is not so smart. If you ask me to send, I will send. Anyway, it is you who ask me to send, and I don’t care about anything else.

• Stateful Services

TCP is a stateful service. Stateful means: I record what was sent, what was not sent, what was received, what was not received, and which should be received. There can be no mistakes. TCP does a lot of things!

UDP is not a stateful service. I just send the message and leave the rest to the receiver. It's a bit arbitrary, isn't it?

How to make UDP realize TCP functions?

As mentioned above, establishing a connection involves three-way handshake and four-way handshake, which UDP can also simulate.

There are a few more questions:

• Sequence Problem
• Packet loss problem
• Flow Control
• Congestion Control

The TCP data structure looks like this:

In fact, if you can explain these structures clearly, you have already understood the core functions of TCP. Below I will still explain the above four questions in plain language.

The order problem and packet loss problem can be solved by using the confirmation and retransmission mechanism. If a packet is received, a confirmation can be made and an ACK can be sent to the sender to tell him that I have received it. If some packets arrive early, they will be cached. If a packet is lost, it can be retried after a timeout. The timeout retry should not be too short, and the time must be greater than the round-trip time RTT, otherwise it will cause unnecessary retransmissions. It should not be too long either. If the timeout is too long, access will become slower. How to determine this time? You can sample the RTT time and perform a weighted average. It also needs to change dynamically according to the network conditions. You can learn about the adaptive retransmission algorithm.

Flow control is to adjust the rate of sending packets according to the network conditions. It uses a sliding window. When confirming a packet, it will also carry a window size. As long as the window size is used well, the rate of sending packets can be adjusted well. The sent message segments should not exceed the window size. (Picture from the Internet)

Congestion control is mainly used to avoid packet loss and timeout retransmission. If these two phenomena occur, it means that the transmission rate is too fast. So how do you know the transmission rate in the beginning? In fact, only one segment of data is sent at the beginning. If an acknowledgment is received, the segment is doubled, and so on. When a timeout retransmission is found, it returns to the situation of sending only one segment. This is slow start, which is not suitable. In fact, there is also a fast retransmission algorithm. Simply put, the congestion window is halved and the subsequent speed is increased linearly. As for how the algorithm is implemented, I will not elaborate here. (Picture source network)

So far, I have explained in plain language the difference between UDP and TCP, as well as what functions UDP lacks and how to make up for it to achieve the functions of TCP. I believe that this answer will make the interviewer feel that there is something to it.

This article is reprinted from the WeChat public account "Wukong Chats about Architecture", which can be followed through the following QR code. To reprint this article, please contact the WeChat public account "Wukong Chats about Architecture".