Network latency, often referred to as network delay, is the time required for data to travel from the sender to the receiver. Network latency is different from network speed or bandwidth or throughput, which is the amount of data that can be transferred per unit time. A network can be both high speed and high latency. For example, a 100 Mbps network speed means the network can transfer 100M bits of data per second, while a 500ms network latency means that the data will need 500 ms to reach the receiver.

One way network latency is difficult to measure. Usually, Ping is used to measure the round-trip time instead. The round-trip time is the sum of the network latencies for the transmit path, the receive path, and the echo request processing time of the device being pinged. Note that the transmit and receive paths are not guaranteed to be the same path, although in practice they are usually the same. Even if they are the same path, the transmit and receive latencies can be different. It is like traffic on a road - it can be congested in one direction and running smoothly in the opposite direction. Some network connections, such as ADSL, can have different bandwidth for different directions.

• Media Access Delay
  Shared media networks are networks in which multiple devices shared the same network media. For example, in Wi-Fi, multiple devices may share the same radio spectrum. If two devices transmit at the same time, they will interfere with each other and corrupt the data. This is called a collision. To avoid collisions, before sending, a device must listen to the channel to make sure no device is using it. If the media is busy, the sender must recheck after a random delay. The random delay is to avoid multiple devices all send at the same time once the media becomes idle, causing collisions.
  
  If the media is still busy after the random delay, the sender can recheck again and again but with an exponentially increasing random delay. It is reasoned that if a media is always busy, it is likely to be congested. The exponentially increasing random delay slows down the transmit rate to help alleviate network congestion.
  
  Even with the above measures, collision is still possible and retransmission may be necessarily, causing additional delay. On a busy Wi-Fi network, it is possible for the Media Access Delay to be hundreds of milliseconds.
  
  In the old days, collisions can also occur on wired Ethernet if the devices are connected using repeater hubs, multi-drop coaxial cables, or if the network port is not full-duplex. Nowadays, wired Ethernet are mostly using full duplex switching hubs and collisions cannot occur.
• Transmission Delay
  Transmission delay, also called store-and-forward delay or packetization delay, is the time required to put a data packet onto the network. This delay occurs because a network device must receive the entire packet before processing it. For example, when a router receives the first byte of a packet, it cannot immediately forward the byte to the next hop. It must wait for the entire packet and verify its checksum first.
  
  This delay can become considerable for network speed below 3 Mbps, such as for some broadband links to the Internet. For example, in the author's city, it is common for ISPs to provide ADSL internet connections with 6 Mbps downstream but only 512 Kbps upstream. In this case, sending a 1500-byte packet (= 12000 bits) would incur 23.4ms of transmission delay.
• Propagation Delay
  Network signals travel at the speed of light. At such high speed, you may think the propagation delay is negligible, but it is not. It takes 100ms for light to travel in an optical fiber to the opposite side of the earth using the shortest path, and in real life, the fiber is not following the shortest path. This shows that propagation delay can be significant for long distance. Even for shorter distances, if the signal happens to go through a satellite, it would like 240 to 280 ms for a single hop to beam the signal up and down.
• Queueing and Processing Delay
  If a network link is not fast enough to transmit all the packets, the packets will queue up inside the network device (such as a switch or router or firewall). Sometimes the CPU of a router or firewall may not be fast enough to process all the packets, and this can also leads to queueing. The queue can grow to thousands of packets long, so queueing delay can be significant and the network can be described as congested. As all devices have limited memory, the queue can become full, in this case the network device can only throw the packets away.
  
  To avoid dropping packets, most network protocols will have some sort of flow control. For example, in TCP, the protocol driver will start with a slow packet rate, then ramp up the rate. If the network starts to lose packets (detectable by using acknowledgement packets from the receiver), it will reduce the packet rate.
  
  For the queueing method, in additional to FIFO (first in, first out), some routers can be configured to assign different priorities to different types of traffic. For example, VOIP (voice over IP), video conference or online gaming traffic can be given higher priority as latency can seriously affect the quality of these services. So it is possible for different applications to experience different network latency.

• Response Time
  Certain applications need fast response time. For example, in internet telephony and video conferencing, even a few hundreds of milliseconds of delay can make communication difficult and inefficient. In online multi-player gaming, even 100ms of delay can be a serious disadvantage.
• Connection Start-up and Ramp-up Time
  Many common protocols, such as TCP, has a connection establishment phase before actual data transfer. The connection establishment requires at least one request and response cycle, that is, one round-trip time. Higher level protocols may also have additional negotiation during connection. For example, in HTTP over SSL (https), the client and server may need to negotiate the encryption method and to exchange and verify digital certificates. A high network latency can increase the connection start-up time by several seconds.
  
  To avoid network congestion, TCP uses "slow start". It may start by transmitting one packet, and wait for an acknowledgement from the receiver to confirm the transfer is successful. It may then transmit two packets and wait for the acknowledgement. If it is successful, it may try four packets. This ramps up the transmit rate exponentially until it reaches a certain threshold, after that it may ramp up linearly. If a packet is lost and retransmission is needed, it may reduce the transmit rate by half and then ramp up the transmit rate linearly.
  
  Note: The above is a simplification. The actual algorithm is much more complicated and different versions of different operating systems can behave differently.
  
  It may take many transmit / acknowledgement cycles for a connection to start up and ramp up to take full advantage of the bandwidth. On a high latency network, the connection and ramp up is slower and may take quite a few seconds.

Many network issues, such as network congestion, affects both network performance and network latency. Network performance is difficult to measure. You may need to transmit large amount of data as fast as possible to saturate the network in order to determine its performance. The test is disruptive - it affects other users of the network. It is also difficult to obtain real-time readings, that is, obtain second by second variation in network performance. On the other hand, measuring real-time network latency is easy with Fast Ping. So network latency is often used as a relative indicator for network trouble-shooting and performance tuning.

As a relative indicator, the network latency of a network path can be compared against that of the same path at a different time. For example, if changing the position of a Wi-Fi antenna reduces network latency of a network path, we can conclude that it improves network performance.