Voltar 
This is a common question for anyone delivering critical applications on public and long distance networks. Which metric should I focus on to maximize the speed of data transfer that support modern applications?
Network performance: throughput is what matters most for user experience
If you put yourself in the shoes of a user connecting to an application, what matters is how fast you see your application appear and how fast you can interact with it. What matters from a user experience:
- how fast you establish your connection to the digital assets that provide the different resources
- how fast you load them.
From a user point of view, network performance is about throughput!
How to measure throughput?
First clarification, many people mix up throughput and bandwidth. Although they are related concepts, they measure two different things.
- First, throughput is the speed at which two devices actually transfer data from one to another.
- Second, bandwidth corresponds to the maximum amount of data that can be transferred on a link.
We use the same unit for both metrics: bits or bytes per second.
While throughput can be measured easily, it is quite hard to measure whether this throughput represents the maximum speed a user can get. Quite often network operations will look at the drivers of throughput to identify potential bottlenecks.
What drives throughput?
In case of network degradation or outage, throughput drops. Monitoring network performance is a must to identify when the network is slow and what is the root cause.
Whichever tools you are using (packet analyzer like Wireshark, SNMP polling like PRTG or Cacti, Traffic loading, active testing) you need indicators that will help you understand whether your users can make the most of the network infrastructure to transfer data.
This article explains 3 key metrics of network performance (latency, throughput and packet loss), how they influence the speed of transfer depending on the protocol used (UDP or TCP).
- Latency is the time required to transmit a packet across a network:
- There are different ways to measure latency: round trip, one way, etc.
- Any element on the path used to transmit data can impact latency: end user device, network links, routers, proxies, local area network (LAN), server,…
- The ultimate limit of latency on large networks is… the speed of light.
If you wish learn more about latency and the different ways to measure it, I recommend you take a look at this article.
- Throughput is the quantity of data being sent/received by unit of time
- Packet loss is the number of packets lost per 100 packets sent by a host
Once we understand each of them right, let’s look at how they interact with each other.
Understand the impact of latency and packet loss on throughput
This can help you understand the mechanisms of network slowdowns.
The protocol used for the communication will impact how things work, so we have to analyze things in a different way for UDP and for TCP.
Measuring Network Performance on UDP
Latency has no impact on throughput on UDP
When using UDP, we assume that all packets sent are received by the other party (transmission control is executed at a different layer, often the application itself).
In theory or for some specific protocols (if no control is undertaken at a different layer; e.g., one-way transmissions), the rate at which packets can be sent by the sender is not impacted by the time required to deliver the packets to the other party (= latency). The sender will send a given number of packets per second, which depends on other factors (application, operating system, resources, …).
As a conclusion, latency has no impact on throughput on UDP.
Measuring network performance on TCP
Latency has a direct impact on throughput on TCP
TCP is a more sophisticated protocol: it involves a transmission control which checks the proper delivery of all packets. This mechanism is called acknowledgment: the receiver responds back with a specific packet or flag to the sender to confirm the proper reception of each packet.
TCP Congestion Window
As an optimization, not all packets will be acknowledged one by one; the sender does not wait for each acknowledgment before sending new packets. Indeed, the number of packets that may be sent before r