RDMA Traffic Class Tagging for RoCE Congestion Tuning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
RDMA workloads on network clusters fail to achieve desired throughput due to suboptimal congestion management protocols in RoCE, leading to inefficient use of network resources in virtualized cloud infrastructure.
Innovation Solution
A framework that performs parametric optimizations for different classes of RDMA traffic by extracting tags from packets to determine traffic classes and processing packets accordingly, using network devices with data processors and non-transitory computer-readable storage media to execute instructions for customized processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If standard RoCE congestion management protocol is used, then network infrastructure is simplified, but RDMA workload throughput fails to achieve desired level
Solution Approach 1:
The patent segments RDMA traffic into different classes (e.g., latency-sensitive, bandwidth-sensitive, background traffic) and applies distinct congestion management parameters to each class. This segmentation allows the system to optimize throughput for specific traffic types without requiring complete redesign of the entire congestion management protocol.
Solution Approach 2:
The patent applies local quality by configuring different congestion management parameters (such as queue depths, ECN markings, DSCP settings) at specific network devices and NICs based on the local traffic class requirements. This enables tailored optimization for different traffic flows while maintaining standard protocol infrastructure.
2Productivity
If customized processing for different traffic classes is implemented, then network throughput is enhanced, but network device complexity increases
Solution Approach 1:
The patent performs preliminary action by classifying RDMA traffic into different classes at the source host machine before transmission. Tags are extracted from packets and traffic classes are determined in advance, allowing network devices to apply pre-configured processing rules without performing complex real-time analysis, thus enhancing throughput while limiting processing complexity.
Solution Approach 2:
The patent utilizes parameter changes by adjusting congestion management parameters (queue depths, ECN markings, DSCP settings) based on traffic class. This allows the system to optimize network throughput through parameter tuning rather than structural changes, maintaining relatively simple device architecture while achieving enhanced performance.
3Reliability
If traffic class-specific optimizations are applied, then performance for bandwidth-sensitive and latency-sensitive traffic is improved, but congestion management protocol complexity increases
Solution Approach 1:
The patent segments traffic performance requirements by creating distinct traffic classes for latency-sensitive, bandwidth-sensitive, and background traffic. Each class receives customized processing parameters tailored to its specific performance needs, ensuring reliable performance for critical traffic types without requiring complex protocol changes for all traffic.
Solution Approach 2:
The patent applies local quality by configuring specific congestion management parameters at network devices and NICs for each traffic class location. This enables the system to guarantee performance reliability for specific traffic types through localized parameter optimization while maintaining standard protocol behavior for other traffic, thereby limiting overall protocol complexity.
Data Source
AI summary
Discussed herein is a framework that provisions for customized processing for different classes of traffic. A network device in a communication path between a source host machine and a destination host machine extracts a tag from a packet received by the network device. The packet originates at a source executing on the source host machine and whose destination is the destination host machine. The tag set by the source and indicative of a first traffic class to be associated with the packet, the first traffic class being selected by the source from a plurality of traffic classes. The network device determines the first traffic class based on the tag extracted from the packet and processes the packet based on the first traffic class.