k2 vs Hysteria2: Congestion Control Comparison

Congestion control is one of the most critical factors determining the performance of a tunnel protocol. Under high packet loss and high latency network conditions, different congestion control strategies produce dramatically different user experiences. This article compares k2 and Hysteria2 across four dimensions.

Background

k2 uses a proprietary adaptive congestion control algorithm. It automatically finds the optimal sending rate, dynamically responding to changing network conditions.

Hysteria2 uses a congestion control strategy called Brutal: the user specifies a maximum bandwidth cap, and the protocol sends at that fixed rate regardless of congestion signals.

---

Dimension 1: Packet Loss Recovery

k2

k2's proprietary adaptive congestion control algorithm uses a censorship-aware mechanism to handle packet loss. When packet loss is detected, the algorithm distinguishes between congestion-induced loss and non-congestion loss (such as active packet dropping by censorship infrastructure):

• Censorship-aware: Distinguishes between congestion-induced and censorship-induced packet drops, avoiding unnecessary rate reduction under active interference.
• High-censorship optimization: Maintains higher sending rates under non-congestion packet loss, keeping throughput stable in censored network environments.
• Probing mechanism: After a rate reduction, periodically attempts higher rates to actively recover available bandwidth.

Hysteria2

The Hysteria2 Brutal strategy ignores all packet loss signals and sends at a fixed rate. This means:

• No distinction between congestion-induced loss and other types of packet drops.
• No rate backoff; loss is handled entirely by retransmission.
• In high-loss networks, this can lead to retransmission storms that further degrade network conditions.

Conclusion: k2's adaptive packet loss handling is significantly more resilient in high-censorship, high-loss environments.

---

Dimension 2: Latency Stability

k2

k2's proprietary algorithm incorporates RTT (round-trip time) as a key signal for rate adjustment. When buffer queues begin to build up (bufferbloat), rising RTT triggers a rate reduction to suppress latency degradation:

• RTT awareness: Real-time monitoring of round-trip time prevents excessive filling of network buffers.
• Rate-based pacing: Data packets are spread evenly over time, avoiding burst traffic that causes queue buildup.

Hysteria2

Hysteria2's fixed-rate sending strategy is insensitive to RTT changes:

• Continuously fills the link at a constant rate, easily causing router buffer buildup (bufferbloat).
• On shared links, latency can fluctuate significantly as other traffic levels change.

Conclusion: k2's RTT-aware mechanism provides superior latency stability compared to Hysteria2's fixed-rate approach.

---

Dimension 3: Bandwidth Utilization

k2

k2's proprietary algorithm continuously probes for the optimal sending rate:

• No manual configuration needed: The algorithm automatically discovers and tracks the actual available network bandwidth.
• Probing mechanism: After rate reductions, periodically attempts higher rates to dynamically recover available throughput.
• Adapts to dynamic networks: Performs better on mobile networks or shared links where bandwidth fluctuates frequently.

Hysteria2

Hysteria2 requires the user to manually specify a target bandwidth:

• Dependent on user configuration: If the configured value is below actual available bandwidth, link utilization is suboptimal; if too high, it continuously causes congestion.
• No automatic probing: No capability to dynamically track bandwidth changes.
• Limited use cases: Works best only when the user can accurately determine the link capacity in advance.

Conclusion: k2's automatic bandwidth probing reduces configuration burden and maintains higher effective utilization in dynamic network environments.

---

Dimension 4: Fairness

k2

k2's rate-based pacing mechanism demonstrates good fairness when coexisting with other traffic:

• TCP coexistence: Rate adapts to avoid actively displacing bandwidth from other connections.
• Multi-flow stability: When multiple tunnel connections share a link, bandwidth is distributed relatively equitably across flows.

Hysteria2

Hysteria2's Brutal strategy has significant fairness problems in multi-flow scenarios:

• Crowds out other traffic: Sends at a fixed rate without yielding to other flows, potentially starving TCP connections.
• Congestion collapse risk: In multi-flow concurrent scenarios, this can trigger congestion collapse, dramatically reducing overall throughput.

Conclusion: k2's rate-adaptive strategy is substantially fairer in multi-flow environments and does not actively disrupt other connections.

---

Summary Table

Dimension	k2 (Proprietary Adaptive Algorithm)	Hysteria2 (Brutal)
Packet Loss Recovery	Censorship-aware; distinguishes congestion vs. non-congestion loss	Ignores loss signals; relies on retransmission
Latency Stability	RTT-aware + rate pacing; suppresses bufferbloat	Fixed rate; prone to queue buildup
Bandwidth Utilization	Automatic optimal rate probing; no manual config needed	Requires manual setting; no dynamic probing
Fairness	Rate-adaptive; coexists peacefully with other flows	Fixed rate; tends to crowd out other traffic

---

FAQ

Isn't Brutal with 1Gbps declared bandwidth faster?

On ideal networks, high declared bandwidth saturates the link quickly. But on real cross-border networks, bandwidth exceeding link capacity becomes retransmissions. Worse, Brutal cannot adapt to peak-hour bandwidth drops — when 100Mbps drops to 30Mbps, 70% of data needs retransmission. k2cc automatically tracks actual available bandwidth, always maintaining optimal throughput.

BBR is also adaptive — how is k2cc different?

BBR has no censorship awareness. Under GFW's 26% probabilistic loss, BBR misinterprets censorship-induced drops as congestion, severely underestimating available bandwidth. k2cc distinguishes congestion from censorship loss, maintaining far higher throughput. See k2cc vs BBR for details.

How are the 14 test scenarios defined?

Based on academic research: T0-T5 cover ideal to hostile standard networks, T6-T8 simulate GFW censorship (including USENIX Security 2023's measured 26% loss rate), C1-C5 model real conditions from China's three major ISPs to Japan. The framework is open source — anyone can verify with identical conditions.

Should I choose k2cc or Brutal?

If your network has censorship interference (e.g., China mainland outbound), choose k2cc — Brutal triggers retransmission storms under censorship loss. If you're on an uncensored LAN with known exact bandwidth, Brutal can work. But k2cc is never worse than Brutal in any scenario, and requires zero configuration.

---

Notes

The comparisons above are based on algorithmic design analysis. Benchmark data will be published separately when conditions permit.