TCP congestion control — slow start, congestion avoidance, fast retransmit, CUBIC

4.3.22 · Coding › Computer Networks

Woh do windows jo ek sender ko limit karti hain

HUM KYA control karte hain: cwnd. KYUN: rwnd receiver ki RAM ko protect karta hai; cwnd network ki queues ko — yeh do alag bottlenecks hain.

Hum cwnd ko MSS (Maximum Segment Size) ki units mein measure karte hain clarity ke liye. Throughput $\approx \frac{\text{cwnd}}{\text{RTT}}$, toh bada cwnd = faster — jab tak loss na ho.

1. Slow Start (SS) — jaldi se ballpark dhundo

HOW yeh badhta hai (first principles se derivation): Hum chahte hain ki har ACK receive hone par cwnd 1 MSS badhe.

• Maano cwnd = $W$ segments. Ek RTT mein hum $W$ segments bhejte hain → $W$ ACKs receive karte hain.
• Har ACK 1 add karta hai → ek RTT ke baad, cwnd $=W+W=2W$.

Toh per-RTT: $W\to 2W$. 1 se shuru karke $n$ RTTs ke baad: $\text{cwnd}(n)=2^n\text{ MSS}$

2. Congestion Avoidance (CA) — limit ke paas tiptoe karo

HOW (derivation): Hum chahte hain RTT mein +1 segment, lekin ACKs poore RTT mein aate rehte hain. cwnd = $W$ segments ke saath, $W$ ACKs per RTT aate hain, aur hum chahte hain ki unka total effect +1 MSS ho: $\text{per ACK: }\text{cwnd}+=\frac{\text{MSS}\times \text{MSS}}{\text{cwnd}}$ $W$ ACKs pe sum karne par: $W\times \frac{{\text{MSS}}^2}{W\cdot \text{MSS}}=\text{MSS}$. ✓ Exactly +1 per RTT.

Toh per-RTT: $\text{cwnd}(t+\text{RTT})=\text{cwnd}(t)+1\text{MSS}$

Yeh AIMD ka AI hai (Additive Increase, Multiplicative Decrease).

3. Loss ke baad react karna — do flavours

WHY halve karo (AIMD mein MD)? Aadha karna ek strong, multiplicative back-off hai jo jaldi congestion relieve karta hai aur, additive increase ke saath combine hokar, ek fair, stable sawtooth deta hai jo converge karta hai (AIMD chart se proven — fairness line).

4. Fast Retransmit — clock ka wait mat karo

Isse ek poora RTO (jo hundreds of ms ho sakta hai) bachta hai, aur pipe full rehti hai.

5. CUBIC — modern default (Linux, kernel 2.6.19 se)

$K$ ki derivation (WHY woh cube root): Loss ke baad, cwnd $\beta W_{\max }$ tak drop hota hai... ruko — actually CUBIC $(1-\beta )$ factor se drop karta hai, toh new window $=(1-\beta )W_{\max }$ hoti hai. Hum chahte hain ki cubic curve $t=0$ par new window se guzre aur $t=K$ par $W_{\max }$ tak pahunche (uska inflection plateau). $W_{\text{cubic}}(0)=(1-\beta )W_{\max }$ set karo: $C(0-K{)}^3+W_{\max }=(1-\beta )W_{\max }$ $-C{K}^3=-\beta W_{\max }\Rightarrow K=\sqrt[3]{\frac{\beta W_{\max }}{C}}$

Worked Examples

Common Mistakes

Flashcards

Connections

• TCP Reliable Data Transfer — cumulative ACKs aur retransmission timers (RTO) jin par yeh build karta hai.
• Flow Control vs Congestion Control — rwnd vs cwnd, do distinct bottlenecks.
• AIMD Fairness — fairness line par convergence ka geometric proof.
• Queueing and Router Buffers — jahan dropped packets actually hote hain (Bufferbloat).
• BBR Congestion Control — loss-based CUBIC ka model-based alternative.
• Bandwidth Delay Product — kyun long fat networks ne CUBIC ko motivate kiya.

Concept Map