Like the road to a car, TCP was the transport to my digital footprint; traffic lights and all. Maybe I couldn’t get rid of my digital DNA, perhaps all I needed to do was delete the digital path known as the transport layer. If voyeurs can’t stream what I did last night, did last night ever happen?
Everything turned out fine the next day. If my friends and co-workers viewed the previous night’s events, they didn’t say anything about it. But that’s what got me thinking about killing off TCP. Born in the 80s, TCP had been the dominant digital communications transport my entire career. Novell had its day in the sun, as did Microsoft, but mostly only on local area networks. Cisco collapsed the backbone into IP packets routed over TCP and it’s been all the TCP/IP stack ever since. For decades.
The problem with TCP is latency. TCP Rate = Maximum Segment Size / Round Trip Time. Round Trip Time (RTT) is latency. Measured in nano or micro seconds on a computer but over the WAN, measured in milliseconds (ms). Latency is mostly distance. The medium matters; air is faster than glass, glass is faster than copper, but theoretically, bits travel at the speed of light, so a satellite hop (round trip) is about a quarter second, or 250ms. Note in the algorithm above that this distance metric is in the denominator of measuring TCP throughput, so the farther the distance, the lower the throughput – by design.
The last significant improvement to the TCP spec was in 1984, where version 4 was developed to mitigate the effects of congestive collapse on the network. Given that rate of innovation, TCP needs to just die. And I think TCP did die this week. It’s being replaced by the Remote Direct Memory Access (RDMA) protocol. The RoCEv2 (pronounced Rocky v2 for RDMA over Converged Ethernet version two) protocol can transport RDMA frames over an IP header and UDP, but Vcinity has a proprietary implementation where they encapsulate the RDMA frames in an IP header for WAN routing and add their own algorithms for flow control and packet loss recovery. The result is an order of magnitude improvement over TCP in throughput.
Think about that. In tech, an order of magnitude improvement generally equates to disruptive technology. A product killer. TCP’s days are numbered.