Skip to content

UDP

UDP is a transport layer protocol. DNS is an application layer protocol that runs on top of UDP (most of the times). Before jumping into UDP, let's try to understand what an application and transport layer is. DNS protocol is used by a DNS client (eg dig) and DNS server (eg named). The transport layer makes sure the DNS request reaches the DNS server process and similarly the response reaches the DNS client process. Multiple processes can run on a system and they can listen on any ports. DNS servers usually listen on port number 53. When a client makes a DNS request, after filling the necessary application payload, it passes the payload to the kernel via sendto system call. The kernel picks a random port number (>1024) as source port number and puts 53 as destination port number and sends the packet to lower layers. When the kernel on server-side receives the packet, it checks the port number and queues the packet to the application buffer of the DNS server process which makes a recvfrom system call and reads the packet. This process by the kernel is called multiplexing (combining packets from multiple applications to same lower layers) and demultiplexing (segregating packets from single lower layer to multiple applications). Multiplexing and Demultiplexing is done by the Transport layer.

UDP is one of the simplest transport layer protocol and it does only multiplexing and demultiplexing. Another common transport layer protocol TCP does a bunch of other things like reliable communication, flow control and congestion control. UDP is designed to be lightweight and handle communications with little overhead. So, it doesn’t do anything beyond multiplexing and demultiplexing. If applications running on top of UDP need any of the features of TCP, they have to implement that in their application.

This example from python wiki covers a sample UDP client and server where “Hello World” is an application payload sent to server listening on port number 5005. The server receives the packet and prints the “Hello World” string from the client.

Applications in SRE role

  1. If the underlying network is slow and the UDP layer is unable to queue packets down to the networking layer, sendto syscall from the application will hang till the kernel finds some of its buffer is freed. This can affect the throughput of the system. Increasing write memory buffer values using sysctl variables net.core.wmem_max and net.core.wmem_default provides some cushion to the application from the slow network
  2. Similarly, if the receiver process is slow in consuming from its buffer, the kernel has to drop packets which it can’t queue due to the buffer being full. Since UDP doesn’t guarantee reliability these dropped packets can cause data loss unless tracked by the application layer. Increasing sysctl variables rmem_default and rmem_max can provide some cushion to slow applications from fast senders.