Why every TCP connection starts at one packet (and doubles until something breaks)

Every fresh TCP connection deliberately under-uses your link, then doubles its send rate every round-trip until a packet drops. That crash isn't a failure mode — it's the entire algorithm. It's how TCP probes for available bandwidth without a measurement instrument.

The mechanism has two phases. First, slow-start: cwnd doubles each RTT (1, 2, 4, 8, 16, 32). When the pipe fills and a packet drops, the algorithm flips to congestion avoidance — cwnd halves, then ticks up by one per round-trip. The dashed line between the two phases is called ssthresh.

The honest limit is what makes this matter for anyone shipping production services. Short HTTP requests finish before cwnd has time to climb. You're paying for a gigabit link and delivering a kilobit, because the request ended before slow-start was done probing.

Three fixes the modern stack uses: the initial window is now 10 packets, not 1. Reused connections carry their cwnd over, so HTTP/2 and keep-alive skip the climb entirely. And BBR replaces loss-based detection with direct bandwidth and RTT estimation, so the sender doesn't have to crash a packet to discover its capacity.

Why this matters at work: when a first API call to a new endpoint feels slow but subsequent calls snap, that's slow-start. When latency budgets get blown by short requests on long-RTT paths, this is usually the cause.

Видео Why every TCP connection starts at one packet (and doubles until something breaks) канала Adam Rosler

Комментарии отсутствуют

Информация о видео

12 мая 2026 г. 6:55:58

00:00:55

Adam Rosler

Правообладателям

Жалоба на материал Недопустимый материал Нарушение авторских прав

Комментарии

Другие видео канала

Why every TCP connection starts at one packet (and doubles until something breaks)

vLLM Serves 24x More Queries On The Same GPU — Here's How PagedAttention Works

Why Llama 3 decodes 8x faster — they removed heads, not added compute (GQA explained)

Single agent workflows are mostly solved. Multi agent workflows are ais future.

Why FlashAttention Made 8K Context Free: The One IO Change That Changed Training

Stop feeding your AI everything at once. Most people think longer prompts mean better output.

How neural networks actually learn — backpropagation in 45 seconds, from miss to math

Your RAG returns same-neighborhood, different-fact answers. Here's why

My CLI tests itself now — every bug becomes a regression test automatically

Drop half your KV cache and the next token is still right — H2O eviction explained

Removing one subtraction from LayerNorm cuts ~30% of training compute

Why your AI agent keeps picking the wrong tool (and how to fix the descriptions)

Google just put a model on your phone. No cloud, no API bill.

AGAR, SUPERKID, KOBE

The internet routes on the honor system — how a 2008 BGP leak killed YouTube globally

DeepSeek's KV Cache Is 1/16 The Size Of Llama's — Here's The Math

ChatGPT isn't thinking. Here's the 6 steps of math that actually happen.

GRPO: how DeepSeek-R1 trained reasoning without a critic, reward model, or human labels

Your 1MB write hits disk as 30MB on a healthy RocksDB instance

Qwen Trained On 32K Tokens, Shipped At 1M — Here's The YaRN Trick

If you're not using Markdown with your AI, you're wasting it.