The Gap Between Default and Reality
Think of a highway designed in the 1950s for 30 mph traffic. It works fine for its era, but put modern vehicles on it and you have artificial bottlenecks everywhere — not because the cars are slow, but because the infrastructure was never designed for them. TCP/IP defaults have the same problem.
Most operating systems ship with network stack parameters tuned for the networks of the early 1990s: 10 Mbps Ethernet, sub-millisecond LAN latency, and machines with 4–16 MB of total RAM. These constraints shaped every default buffer size, queue depth, and timer value baked into the kernel. Today, those defaults are still present — even on a freshly provisioned cloud instance with a 10 Gbps NIC.
The practical consequence is severe. A default Linux server transferring data over a trans-Pacific link (say, Virginia to Tokyo, ~170ms RTT) on a 1 Gbps connection may achieve only 2–5 Mbps of actual throughput. The wire is fast. The NIC is fast. The bottleneck is a 64 KB buffer that was sized for a world where 64 KB was generous. TCP/IP performance tuning is the discipline of adjusting kernel parameters, queue disciplines, and protocol options to match the physics of the actual network rather than the assumptions of a bygone era.
The parameters you tune fall into three broad categories: buffer sizing (how much data can be in flight), congestion control (how the sender responds to network signals), and queue management (how routers and hosts handle bursts). Each category addresses a different failure mode, and understanding all three is necessary to diagnose real-world throughput problems.
A common interview question is: "A customer complains their 1 Gbps link to a remote data center only achieves 5 Mbps. Where do you start?" The correct answer begins with calculating the Bandwidth-Delay Product (BDP) and comparing it to the system's TCP buffer limits — not with checking the NIC, the firewall rules, or the application. Demonstrating that you know the bottleneck is mathematical before it is physical separates senior candidates from mid-level ones.
Key Point: Default TCP/IP settings are optimized for 1990s LAN conditions. On modern high-speed, high-latency links, they create artificial throughput ceilings that have nothing to do with the physical capacity of the network.