Time Difference Congestion Notification for Data Center Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current congestion control techniques in data center networks fail to provide high throughput and ultra-low latency with low computational overhead, as they are either slow in response, unable to accurately detect congestion, or incur high maintenance costs, especially as data center networks grow in size and complexity.
Innovation Solution
The Time Difference Congestion Notification (TDCN) system uses a three-point architecture with hardware-generated timestamps at the source and destination devices to calculate a one-way travel time difference, allowing for precise congestion control by adjusting packet transmission rates based on congestion levels, thereby reducing computational overhead and ensuring reliable data transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional TCP based congestion control is used, then packet loss detection is achieved, but response time is slow and throughput is limited
Solution Approach 1:
The patent applies preliminary action by using ECN flags set by switches to notify congestion before packet loss occurs. The source device receives congestion notifications in advance through acknowledgment packets, allowing it to proactively reduce transmission rate before buffer overflow and packet loss happen, thus improving response time while maintaining reliability
Solution Approach 2:
The patent implements feedback mechanism where switches set ECN flags in data packets when congestion is detected, and destination devices return congestion notifications in acknowledgment packets to the source. This closed-loop feedback enables the source to continuously adjust transmission rate based on real-time network conditions, resolving the contradiction between fast response and reliable delivery
2Measurement precision
If ECN based congestion control is used, then congestion detection is achieved, but congestion extent detection is not possible
Solution Approach 1:
The patent applies partial action by using ECN flags not just for binary congestion notification but for quantized congestion measurement. Different ECN flag patterns and multiple congestion notification packets provide partial information about congestion extent, allowing the source to estimate queue length and adjust rate proportionally rather than with simple on/off control
Solution Approach 2:
The patent changes the parameter interpretation of ECN flags from simple presence/absence to quantized congestion levels. By analyzing the frequency and pattern of ECN-flagged packets and congestion notifications, the source device infers the extent of congestion (queue length) and adjusts transmission rate accordingly, recovering lost information through parameter transformation
3Measurement precision
If DCTCP congestion control is used, then congestion measurement is improved, but computational overhead and latency increase
Solution Approach 1:
The patent applies self-service by having switches automatically set ECN flags based on their own queue length measurements, without requiring complex host-based congestion detection algorithms. The destination device also performs self-service by automatically generating congestion notifications in acknowledgment packets, eliminating the need for complex middlebox infrastructure and reducing overall system complexity while maintaining measurement accuracy
4Device complexity
If DCQCN congestion control is used, then lightweight implementation is achieved, but performance deteriorates in large networks
Solution Approach 1:
The patent applies segmentation by dividing the congestion control function into three independent components: switches that set ECN flags based on local queue conditions, destination devices that generate congestion notifications, and source devices that adjust transmission rates. This segmentation allows each component to operate independently and efficiently, enabling fast convergence even in large networks while maintaining implementation simplicity
Solution Approach 2:
The patent uses preliminary action by having switches set ECN flags immediately when queue length exceeds thresholds, providing early congestion warning before significant packet loss occurs. This preliminary notification allows the source device to proactively reduce rate, achieving fast convergence without waiting for packet loss or RTT measurements, thus maintaining high speed performance in large networks
Data Source
AI summary
A system is provided for controlling congestion. The system sends a first data packet to a destination device, and receives a first acknowledgement packet from the destination device corresponding to the first data packet. Next, the system can extract, from the first acknowledgement packet, a first pair of timestamp values which includes a first time stamp value generated by the source device and a second timestamp value generated by the destination device, and further extract, from a second acknowledgement packet corresponding to a subsequently transmitted packet to the destination device, a second pair of time stamp values which include a third time stamp value generated by the source device and a fourth time stamp value generated by the destination device. The system can determine the congestion change by calculating a one-way travel time difference based on the first pair of timestamp values with the second pair of timestamp values.