tcp congestion control algorithms

overview

tcp congestion control prevents network collapse by adapting sending rate to network conditions. it operates independently from flow control, using packet loss and delay as signals to adjust the congestion window (cwnd).

congestion window dynamics

slow start phase

congestion avoidance (aimd)

fast retransmit and recovery

congestion window evolution

algorithm comparison

cubic algorithm

cubic window function

W(t) = C * (t - K)³ + W_max

Where:
- t = time since last reduction
- K = ∛(W_max * β / C)
- β = multiplicative decrease factor (0.7)
- C = scaling constant (0.4)
- W_max = window size before last reduction

bbr (bottleneck bandwidth and rtt)

congestion signals

loss-based signals

delay-based signals

RTT_min = minimum observed RTT (empty queue)
RTT_current = current RTT measurement
Queue_delay = RTT_current - RTT_min

If Queue_delay > threshold:
    Reduce sending rate
Else:
    Increase sending rate

implementation details

key variables

struct tcp_congestion {
    uint32_t cwnd;           // Congestion window
    uint32_t ssthresh;       // Slow start threshold
    uint32_t flight_size;    // Bytes in flight
    uint32_t recovery_point; // Seq at fast recovery entry
    uint8_t  ca_state;       // Congestion state
    uint32_t dup_ack_count;  // Duplicate ACK counter

    // Algorithm-specific
    union {
        struct cubic_state cubic;
        struct bbr_state bbr;
        struct vegas_state vegas;
    } algo_state;
};

linux congestion control

# View available algorithms
sysctl net.ipv4.tcp_available_congestion_control
# cubic reno bbr

# Current algorithm
sysctl net.ipv4.tcp_congestion_control
# cubic

# Change algorithm
sudo sysctl -w net.ipv4.tcp_congestion_control=bbr

# Per-socket algorithm
setsockopt(sock, IPPROTO_TCP, TCP_CONGESTION, "bbr", 4);

performance characteristics

tuning parameters

initial window (iw)

# RFC 6928 recommends IW = 10 MSS
# Linux default (segments)
sysctl net.ipv4.tcp_init_cwnd

slow start after idle

# Restart slow start after idle period
sysctl net.ipv4.tcp_slow_start_after_idle
# 1 = enabled (default), 0 = disabled

ecn (explicit congestion notification)

# ECN negotiation
sysctl net.ipv4.tcp_ecn
# 0 = disabled
# 1 = enabled when requested
# 2 = enabled by default

common issues

throughput collapse

symptoms:

• repeated timeouts
• very low throughput
• cwnd stuck at minimum

causes:

• severe packet loss
• incorrect rto calculation
• middlebox interference

unfairness

symptoms:

• one flow dominates
• starvation of competing flows
• unequal bandwidth sharing

causes:

• rtt unfairness (reno)
• aggressive algorithms
• different algorithms competing

bufferbloat

symptoms:

• high latency under load
• variable performance
• poor interactive response

solutions:

• use bbr or vegas
• enable ecn
• implement aqm at router

monitoring and debugging

# Monitor congestion metrics
ss -tin | grep -E "cwnd|ssthresh|rtt"

# Track retransmissions
netstat -s | grep -i retrans

# Watch algorithm behavior
tcpprobe -p 8888 > tcp_probe.log

# BPF trace congestion events
bpftrace -e 'kprobe:tcp_enter_loss { printf("Loss event\\n"); }'

key takeaways

• congestion control prevents network collapse through adaptive rate control
• operates independently from flow control (network vs receiver limitation)
• slow start provides rapid initial bandwidth discovery
• aimd provides fairness and stability
• modern algorithms (cubic, bbr) improve performance in challenging networks
• proper algorithm selection depends on network characteristics

references

• rfc 5681 - tcp congestion control
• rfc 6582 - newreno modification
• rfc 8312 - cubic for fast networks
• rfc 6928 - increasing initial window
• google bbr - congestion-based control

next steps

• review sliding window protocol for flow control
• explore tcp fundamentals for base concepts
• analyze connection lifecycle for debugging