Packets Are the Past, Flows Are the Future
Networking has focused for years on speeds and feeds: how many packets you can handle and how many ports per device. But hosts and applications work with flows, which will force the network to adapt. Here’s why.
November 14, 2012
Speeds and feeds have been the networking story for the last two decades. Vendors have battled to see which device can push out the most packets per second, or pare down Ethernet frame latency to the fewest milliseconds, or offer up the biggest MAC tables. And as Ethernet speeds increased, so did feeds: 100-Mbps Ethernet is now 10 Gbps, while port counts have jumped from tens to hundreds per chassis.
However, networking is changing because the future is about flows. Packet and ports are no longer the metrics by which the network will be judged. Application clients and servers aren't aware of packets or ports. The OS opens a session with a remote host and makes the connection via the transport layer using TCP, and dispatches a stream of data over the network. (That's why TCP is a definition for propagating a stream of data.)
That stream, or flow, is run through the IP process to segment the stream into a consistent packet data structure, which is easily routed through network paths. Packets enable resilient network designs by creating uniform, stateless, independent data chunks. Packet networking has been the preferred method for networking for the last 20 years (but it's not only one). The use of packets remains a sound technical concept for abstracting the forwarding path from the payload: Focus on the network, don't be concerned about the data.
But consider common use cases for the network. The user wants to see a Web page load smoothly and quickly. The Web developer wants the Web browser to request an image, an HTML file or a script file from the Web server and load rapidly. For a sys admin, the OS opens a socket to manage a connection to a remote system and to stream data through it. These use cases are all about flows of data, but the network supports the delivery as a series of packets.
And consider the hosts at either end of the connection. In the past, servers were always located on a single port of a switch. Clients were located at a specific WAN connection or on a specific switch in the campus LAN. Packets ingressed and egressed the network at known locations--the ports.
The time of packets and ports is passing. In virtualized server environments, the network ingress for a given application isn't physical, can change regularly, and is configured by the VM administrator. Users are moving from desktops to handsets and laptops. The network must now deliver data flows from server to client. That's a fundamental change in network services. Clearly, networking by packets doesn't match the use cases, yet today's network engineer is concerned only with packets and ports.
Communication between application clients and servers has always been about flows, not packets, but the technical challenges of configuring the network have hidden that issue behind technical abstraction of packets. As technology has improved, packet forwarding has become commoditized. The port density and performance of the current generation of network devices is good enough for majority of the market.
And that's why emerging technologies such as software-defined networking (SDN) and OpenFlow are gaining acceptance. They are designed to enable applications and data flows, not just to move packets from point A to point B. In a subsequent post, I'll look at how SDN is a business technology while OpenFlow is a networking technology.
About the Author
You May Also Like