Flow-Aware Congestion Window Control for Data Center Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Data Center Networks (DCNs) experience congestion due to network packet traffic, leading to inefficiencies in data transfer and latency, and existing congestion control protocols struggle to adaptively manage packet transmission rates based on varying network conditions and flow dynamics.
Innovation Solution
A congestion window adjustment mechanism that dynamically adjusts the transmission rate of packets based on real-time network congestion levels and the number of active flows contributing to congestion, using mechanisms like Explicit Congestion Notification (ECN) and Priority-Flow-Control (PFC) to optimize packet injection rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If congestion control protocols adjust packet transmission rates based on network congestion information, then network stability and bandwidth utilization improve, but latency increases due to the overhead of congestion detection and adaptation
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing congestion window sizes in a lookup table based on ECN marks and PFC queue depths. When congestion is detected, the sender can immediately retrieve the pre-computed CWND value without performing real-time calculations, thus maintaining network stability while minimizing the time overhead for congestion adaptation.
Solution Approach 2:
The patent introduces an intermediary mechanism by using PFC pause queues and ECN marks as intermediate indicators of network congestion. Instead of directly measuring complex network conditions, the system uses these intermediary signals (pause queue depth, ECN mark frequency) to trigger congestion window adjustments, simplifying the detection process and reducing latency.
2Productivity
If the congestion window size is increased to improve throughput, then bandwidth utilization improves, but network congestion worsens leading to packet loss and retransmissions
Solution Approach 1:
The patent implements feedback by continuously monitoring ECN marks and PFC pause queue depths, and using these measurements to dynamically adjust the congestion window size. The system retrieves pre-calculated CWND values from lookup tables based on current congestion indicators, creating a closed-loop control mechanism that increases throughput while preventing excessive congestion through real-time adaptation.
Solution Approach 2:
The patent applies parameter changes by dynamically modifying the congestion window size parameter based on network conditions. The lookup tables store CWND values corresponding to different levels of ECN marks and PFC queue depths, allowing the system to change the transmission rate parameter adaptively to maximize throughput while avoiding network overload.
3Productivity
If multiple flows share the network simultaneously, then network resource utilization improves, but congestion control complexity increases making it difficult to maintain fairness
Solution Approach 1:
The patent applies local quality by implementing flow-specific congestion control parameters. Each flow maintains its own congestion window and uses the lookup table based on its specific ECN mark history and PFC queue depth observations. This allows multiple flows to share network resources efficiently while each flow's congestion control behavior is independently optimized, maintaining fairness without requiring complex global coordination.
Data Source
AI summary
Examples described herein relate to a network interface device that is to adjust a transmission rate of packets based on a number of flows contributing to congestion and/or based on whether latency is increasing or decreasing. In some examples, adjusting the transmission rate of packets based on a number of flows contributing to congestion comprises adjust an additive increase (AI) parameter based on the number of flows contributing to congestion. In some examples, latency is based on a measured roundtrip time and a baseline roundtrip time.