Switch-Assisted Transport for Faster Datacenter Bandwidth Recovery

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing congestion protocols in datacenters, such as TCP, suffer from slow recovery of available bandwidth, leading to detrimental effects on network performance, throughput, latency, resource utilization, and fairness, particularly in applications like Distributed Machine Learning and Big Data, which require scalable, stateless, and application-aware optimal transport protocols.

Innovation Solution

The Omni-Transport Mechanism (Omni-TM) employs a switch-assisted transport layer with network-aware rate adjustment using exponentially weighted moving averages to determine aggregated demand values without maintaining per-flow states, allowing for improved link and resource utilization by coordinating between network devices and applications in a distributed and stateless manner.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If TCP congestion protocol is used, then network reliability is maintained, but bandwidth recovery speed is slow

Engineering Contradiction:
Improvenetwork reliabilityVSAvoidbandwidth recovery speed
Core Design Contradiction:
ReliabilityVSSpeed

Solution Approach 1:

The patent changes the control parameters of congestion response by introducing ECN-marked packets as a signaling mechanism. Instead of traditional TCP slow start recovery, the system uses ECN marks to trigger faster rate adjustments at the application layer, fundamentally altering how congestion information is processed and acted upon to accelerate bandwidth recovery while maintaining reliability

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces an intermediary mechanism between the network layer and transport layer by using ECN-marked packets as carriers of congestion information. These marked packets serve as intermediaries that convey network status without requiring full TCP retransmission handshakes, enabling faster response while preserving the reliability guarantees of TCP

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If stateful congestion control mechanisms are implemented, then application performance is improved, but device complexity increases

Engineering Contradiction:
Improveapplication performanceVSAvoiddevice complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent extracts the congestion control state maintenance from the network device to the application layer. Instead of requiring switches/routers to maintain per-flow state, the system uses ECN marks that carry aggregated congestion information, removing the complex stateful mechanisms from network devices while preserving performance benefits through application-aware rate adjustment

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If network-aware rate adjustment is implemented, then link utilization is improved, but protocol complexity increases

Engineering Contradiction:
Improvelink utilizationVSAvoidprotocol complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent makes ECN marks serve multiple functions: they simultaneously convey congestion information, trigger rate adjustment, and enable application-aware optimization. This multi-functionality allows the system to achieve improved link utilization without adding dedicated complex protocol mechanisms, as the existing ECN infrastructure is leveraged for multiple purposes

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentUS12463905B2System and methods for network-aware rate adjustment
Publication Date: 2025.11.04 HUAWEI TECH CO LTD
  • US12463905B2 patent drawing
  • US12463905B2 patent drawing
  • US12463905B2 patent drawing

AI summary

The disclosure may provide methods, systems and apparatus related to a switch-assisted transport layer using network-aware rate adjustment. According to an aspect, a method by a network device may be provided. The method includes receiving one or more data packets corresponding to one or more flows to be forwarded through an egress port of the network device. Each data packet may indicate a flow demand per time period T for a corresponding flow of the one or more flows. The method further includes determining a demand value for the egress port based on the indicated flow demand. The method further includes sending the one or more data packets through the egress port. Each data packet sent may indicate a value for an aggregated demand of the corresponding flow, The value for the aggregated demand may be based on the determined demand value.