In this performance evaluation, we tested three congestion control algorithms—NDM-TCP (ML-based), TCP Cubic, and TCP Reno—under near-optimal network conditions that simulate high-quality fiber or broadband connections.
Test Configuration: Optimal Fiber/Broadband Simulation
To replicate high-performance wired network conditions, we configured the following network parameters using Linux traffic control (tc):
- Base Latency: 15ms
- Latency Variation: ±2ms (minimal jitter)
- Packet Loss Rate: 0.1% (near-perfect reliability)
These parameters were randomly chosen to represent ideal conditions in modern fiber-optic or high-quality broadband networks, where users experience low latency, minimal jitter, and extremely low packet loss.
Test Results Overview
The following data was captured during a 10-second iperf3 performance test:
| Metric | NDM-TCP (ML Model) | TCP Cubic | TCP Reno |
|---|---|---|---|
| Total Transfer (Sender) | 838 MBytes | 825 MBytes | 740 MBytes |
| Total Received (Receiver) | 835 MBytes | 824 MBytes | 738 MBytes |
| Average Bitrate (Sender) | 702 Mbits/sec | 692 Mbits/sec | 620 Mbits/sec |
| Receiver Bitrate | 698 Mbits/sec | 689 Mbits/sec | 617 Mbits/sec |
| Total Retransmissions | 10 | 20 | 22 |
| Test Duration (Receiver) | 10.03 sec | 10.03 sec | 10.03 sec |
Detailed Interval Analysis
NDM-TCP Performance Over Time
| Interval (sec) | Transfer | Bitrate | Retr | Cwnd |
|---|---|---|---|---|
| 0.00-1.00 | 99.4 MBytes | 833 Mbits/sec | 2 | 21.4 MBytes |
| 1.00-2.00 | 116 MBytes | 975 Mbits/sec | 1 | 7.31 MBytes |
| 2.00-3.00 | 117 MBytes | 984 Mbits/sec | 0 | 7.81 MBytes |
| 3.00-4.00 | 102 MBytes | 853 Mbits/sec | 1 | 3.31 MBytes |
| 4.00-5.00 | 60.5 MBytes | 508 Mbits/sec | 1 | 1.94 MBytes |
| 5.00-6.00 | 81.8 MBytes | 686 Mbits/sec | 0 | 3.06 MBytes |
| 6.00-7.00 | 86.6 MBytes | 727 Mbits/sec | 2 | 1.50 MBytes |
| 7.00-8.00 | 68.1 MBytes | 571 Mbits/sec | 0 | 2.62 MBytes |
| 8.00-9.00 | 61.1 MBytes | 513 Mbits/sec | 1 | 1.62 MBytes |
| 9.00-10.00 | 43.6 MBytes | 365 Mbits/sec | 2 | 1.19 MBytes |
TCP Cubic Performance Over Time
| Interval (sec) | Transfer | Bitrate | Retr | Cwnd |
|---|---|---|---|---|
| 0.00-1.00 | 95.8 MBytes | 802 Mbits/sec | 3 | 5.81 MBytes |
| 1.00-2.00 | 94.2 MBytes | 790 Mbits/sec | 4 | 2.37 MBytes |
| 2.00-3.00 | 67.4 MBytes | 565 Mbits/sec | 4 | 1.62 MBytes |
| 3.00-4.00 | 59.1 MBytes | 496 Mbits/sec | 1 | 1.94 MBytes |
| 4.00-5.00 | 80.9 MBytes | 679 Mbits/sec | 0 | 3.00 MBytes |
| 5.00-6.00 | 88.9 MBytes | 746 Mbits/sec | 1 | 2.94 MBytes |
| 6.00-7.00 | 108 MBytes | 905 Mbits/sec | 0 | 4.00 MBytes |
| 7.00-8.00 | 103 MBytes | 863 Mbits/sec | 1 | 3.62 MBytes |
| 8.00-9.00 | 61.4 MBytes | 515 Mbits/sec | 3 | 1.75 MBytes |
| 9.00-10.00 | 66.5 MBytes | 557 Mbits/sec | 3 | 1.94 MBytes |
TCP Reno Performance Over Time
| Interval (sec) | Transfer | Bitrate | Retr | Cwnd |
|---|---|---|---|---|
| 0.00-1.00 | 88.8 MBytes | 743 Mbits/sec | 4 | 1.56 MBytes |
| 1.00-2.00 | 74.1 MBytes | 622 Mbits/sec | 3 | 1.31 MBytes |
| 2.00-3.00 | 48.8 MBytes | 409 Mbits/sec | 2 | 2.31 MBytes |
| 3.00-4.00 | 75.2 MBytes | 631 Mbits/sec | 1 | 2.56 MBytes |
| 4.00-5.00 | 56.2 MBytes | 472 Mbits/sec | 3 | 1.75 MBytes |
| 5.00-6.00 | 79.9 MBytes | 670 Mbits/sec | 2 | 1.87 MBytes |
| 6.00-7.00 | 79.9 MBytes | 670 Mbits/sec | 1 | 2.00 MBytes |
| 7.00-8.00 | 85.6 MBytes | 718 Mbits/sec | 2 | 2.00 MBytes |
| 8.00-9.00 | 93.6 MBytes | 785 Mbits/sec | 0 | 4.00 MBytes |
| 9.00-10.00 | 57.4 MBytes | 481 Mbits/sec | 4 | 1.31 MBytes |
Key Findings and Analysis
1. NDM-TCP: Aggressive and Efficient Leader
In this high-quality network environment, the ML-driven NDM-TCP demonstrated optimal performance:
- Highest Throughput: Achieved 702 Mbits/sec, outperforming both Cubic and Reno
- Best Efficiency: Only 10 retransmissions—half that of Cubic and less than half of Reno
- Massive Windows: Reached peak congestion windows up to 21.4 MBytes, demonstrating the ML model's ability to exploit high-bandwidth, low-loss conditions
- Peak Performance: Hit 984 Mbits/sec in the 2-3 second interval with zero retransmissions
2. TCP Cubic: Close Second with More Overhead
TCP Cubic performed competitively but couldn't match NDM-TCP's efficiency:
- Strong Throughput: Achieved 692 Mbits/sec, just 1.4% behind NDM-TCP
- Double the Retransmissions: 20 retransmissions versus NDM-TCP's 10
- Moderate Windows: Peak congestion window of 5.81 MBytes—significantly smaller than NDM-TCP's aggressive scaling
- Consistent Mid-Range Performance: Maintained steady performance in the 500-900 Mbits/sec range
3. TCP Reno: Falling Behind in High-Speed Networks
TCP Reno showed its age in this high-performance scenario:
- Lowest Throughput: Achieved only 620 Mbits/sec, 13% behind NDM-TCP
- High Retransmissions: 22 retransmissions, the worst among all three
- Conservative Windows: Maximum congestion window of 4.00 MBytes
- High Variability: Performance ranged from 409 to 785 Mbits/sec with significant fluctuations
Performance Comparison
Throughput Rankings
- NDM-TCP: 702 Mbits/sec (+1.4% vs Cubic, +13% vs Reno)
- TCP Cubic: 692 Mbits/sec (+12% vs Reno)
- TCP Reno: 620 Mbits/sec
Reliability Rankings (Lower is Better)
- NDM-TCP: 10 retransmissions (Best efficiency)
- TCP Cubic: 20 retransmissions (50% more than NDM-TCP)
- TCP Reno: 22 retransmissions (120% more than NDM-TCP)
Efficiency Metric: Data per Retransmission
- NDM-TCP: 83.8 MBytes per retransmission (Most efficient)
- TCP Cubic: 41.2 MBytes per retransmission
- TCP Reno: 33.6 MBytes per retransmission (Least efficient)
Peak Congestion Window Comparison
- NDM-TCP: 21.4 MBytes (Largest, most aggressive scaling)
- TCP Cubic: 5.81 MBytes
- TCP Reno: 4.00 MBytes (Smallest, most conservative)
Conclusion: ML Advantage in High-Performance Networks
In optimal network conditions with minimal latency and packet loss, NDM-TCP demonstrates a clear advantage:
NDM-TCP excels in high-quality networks by aggressively scaling its congestion window to exploit available bandwidth while maintaining superior efficiency. The ML model's ability to recognize favorable network conditions allowed it to achieve:
- Highest throughput (702 Mbits/sec)
- Lowest retransmissions (10 total)
- Most aggressive window scaling (21.4 MBytes peak)
- Best data efficiency (83.8 MBytes per retransmission)
TCP Cubic remains competitive with 692 Mbits/sec throughput, making it a solid general-purpose choice. However, its doubled retransmission count indicates it's less efficient at exploiting optimal conditions.
TCP Reno struggles in high-bandwidth environments, constrained by its conservative congestion control designed for earlier network generations. Its 13% lower throughput and higher retransmission rate make it suboptimal for modern high-speed networks.
Key Insight: This test reveals that NDM-TCP's machine learning approach isn't just about conservative stability—it's about network awareness. In poor conditions, it prioritizes reliability; in excellent conditions, it aggressively maximizes throughput while maintaining superior efficiency. This adaptive intelligence gives it an edge across diverse network environments.
Top comments (0)