Network performance metrics pose data center challenges

The networking world is known for widespread changes in mindset and often drives the hype cycle in the networking industry. There are many paradigm shifts such as centralized control platforms, decentralized control platforms, and various routing protocols and paradigms.

[[251486]]

One major change that is happening now is the leaf-spine data center structure that supports public or private clouds, which can be thought of as selling network, compute and storage resources as units of computing power.

Data center architectures are often based on the CLOS architecture originally designed for telephone networks, where admission control is the primary means of quality of service. In the circuit-switched world, network performance metrics were primarily about allocating a fixed amount of bandwidth and refusing to allow a connection if there was no available bandwidth.

These initial LOS networks have been adapted to computer networks by folding (or passing traffic in both directions) and building different variants such as Benes and Butter. These more general leaf-spine networks present challenges to network performance metrics beyond admittance control.

Here are some of the challenges facing the network.

For example, in a network, there are 256 paths between host 1 and host 2. What kind of network performance metrics can be collected to understand how the network is performing?

Collecting queue depths, drops, and TCP retransmissions for each interface would provide a solid general view of network performance. However, beyond this abstract view, it is difficult to see how to collect information that would help any one application operate more efficiently.

Here's a more concrete example: If an application is performing poorly, and you suspect the problem is in the network, where would you start troubleshooting? It is difficult to trace the path of any packet or packet flow through these types of wide equal-cost multi-path networks to determine where the problem might be.

One possible answer is to add more state to the network, especially in network performance metrics. For example, if you use the IPv6 version of segment routing (SRv6), you can add a header to each packet accepted by the network and add a list of paths to the packet header.

Because SRv6 does not remove or modify this header during switching, examining the packet header at any point in the path will show the path the packet has taken through the network. There are many different ways to add the type of information needed to track individual flows through the network, but each method also involves another trade-off.

Adding these network performance metrics can create a flood of new information that network management systems must consume, manage, categorize, and account for. Humans must also consume this flood of information in the process of troubleshooting issues, capacity planning, and generally understanding the health of the network. These problems have not yet been fully solved, but they may be new directions for research, machine learning, and more powerful architectural approaches in network management system design and deployment.