Chapter 2 - The Application Layer

Reading: Kurose & Ross, chapter 2

2.1 Principles of network applications

Architectures:

• Client-server. An always-on server with a permanent address; clients talk to the server, never to each other. Scaling means adding servers, which is why data centers exist.
• Peer-to-peer. Peers talk directly. Self-scaling, because each new peer brings capacity as well as demand, but harder to manage and secure.

Processes and sockets. A process is a program running on a host. A socket is the door between the application and the transport layer. The application controls everything above the door; the OS controls everything below it, except for the transport protocol choice and a few parameters.

Addressing a process requires two things: the host's IP address, and a port number identifying the process on that host.

What a transport protocol can offer an application:

Service	TCP	UDP
Reliable data transfer	yes	no
Throughput guarantee	no	no
Timing guarantee	no	no
Security	no (TLS adds it)	no
Congestion control	yes	no

Neither offers timing or throughput guarantees. The Internet does not do quality of service by default, and applications compensate.

2.2 The Web and HTTP

HTTP is stateless: the server keeps nothing about previous requests.

Non-persistent HTTP opens a new TCP connection per object. Cost per object: two RTTs plus transmission time, one RTT for the handshake and one for the request and response.

Persistent HTTP reuses the connection, so subsequent objects cost one RTT. This is the default in HTTP/1.1.

Message format. Request line, header lines, blank line, body. Learn the shape well enough to write one by hand, because you will write SMTP messages by hand in P1, and SMTP uses the same shape.

Cookies restore state to a stateless protocol: a response header sets an ID, the browser stores it, and later requests carry it back.

Web caching puts a proxy between the client and the origin. Conditional GET (If-Modified-Since) lets the cache verify freshness without transferring the object again.

HTTP/2 adds framing and multiplexing, so one large object no longer blocks the small ones behind it. It still runs over one TCP connection, so a single lost segment stalls every stream. HTTP/3 over QUIC removes that last bit of head-of-line blocking.

2.3 Electronic mail

Three components: user agents, mail servers, and SMTP.

SMTP is a push protocol: the sender's server pushes to the recipient's server. Contrast HTTP, which is a pull protocol. That asymmetry explains most of the structural differences between them.

SMTP uses persistent connections, requires the message to be 7-bit ASCII, and terminates the body with a lone period on its own line.

Mail access is a separate problem, solved by IMAP or HTTP, because the recipient is not always online to receive a push.

TIP

P1 is an SMTP client. Read RFC 5321 alongside this section, and read section 2.3 before you start writing code.

2.4 DNS

DNS translates hostnames into IP addresses. It is a distributed, hierarchical database, and it is a protocol for querying that database.

The hierarchy: root servers, top-level domain servers, authoritative servers. Plus local DNS servers, which are not strictly in the hierarchy but do most of the work in practice.

Iterative queries: the contacted server replies with the name of the next server to ask. Recursive queries: the contacted server does the work and returns the answer. In practice the host-to-local-server query is recursive and the rest are iterative.

Caching is what makes DNS fast. Once a local server learns a mapping, it caches it, which is why the root servers are not overwhelmed. TTLs bound the staleness.

Resource records: A, NS, CNAME, MX.

2.5 P2P file distribution

The key insight: with client-server, distribution time grows linearly with the number of clients, because the server's uplink is the bottleneck. With P2P it grows far more slowly, because every peer contributes upload capacity.

BitTorrent: chunks, the tracker, the swarm, rarest-first requesting, and the tit-for-tat unchoking that provides the incentive not to freeload.

2.6 Video streaming and CDNs

DASH: the video is encoded at multiple rates, chopped into chunks, and the client picks the rate per chunk based on the bandwidth it is currently measuring. The intelligence sits in the client.

CDNs push content to the network edge, and use DNS redirection to steer each client to a nearby replica.

2.7 Socket programming

The call sequence for TCP:

• Server: socket, bind, listen, accept, then read/write on the returned socket.
• Client: socket, connect, then read/write.

For UDP there is no connection: socket, bind on the server, then sendto and recvfrom on both sides.