Virtual Routing Fields for Isolating Incast Network Congestion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-performance computing systems face challenges in managing network congestion due to the incast problem, where multiple nodes send data simultaneously to fewer nodes, leading to severe traffic overload and inefficient resource allocation, which existing Explicit Congestion Notification (ECN) techniques respond to only gradually, affecting performance.
Innovation Solution
Implementing virtual routing fields (VRFs) in switches to augment routing decisions by configuring isolation between different traffic types and enabling load balancing, allowing for dynamic adjustment of network paths without altering existing routing algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If adaptive routing is used in congested networks, then routing flexibility and resource utilization are improved, but congestion spreads to more links when receivers are overloaded with incast traffic
Solution Approach 1:
The network fabric is segmented into multiple virtual routing fields (VRFs) that can be independently controlled. Incast-prone traffic is isolated into dedicated VRFs separate from latency-sensitive traffic, preventing congestion spread while maintaining routing flexibility within each segment.
Solution Approach 2:
Virtual routing fields act as intermediary layers between physical network links and traffic flows. These VRFs mediate traffic routing decisions, allowing adaptive routing algorithms to operate within isolated virtual spaces without propagating congestion to other traffic types.
2Productivity
If multiple compute nodes send high-volume traffic to few storage nodes simultaneously, then data transfer throughput is improved, but network infrastructure becomes overloaded causing incast problem
Solution Approach 1:
Storage traffic is segmented into dedicated virtual routing fields that provide isolated pathways from compute nodes to storage nodes. This segmentation allows high-volume data transfers to proceed reliably without overloading shared network infrastructure, as each VRF handles specific traffic patterns independently.
Solution Approach 2:
The system dynamically assigns traffic to different virtual routing fields based on traffic type and destination. Compute nodes can dynamically switch between VRFs for different storage operations, enabling flexible high-throughput data transfer while maintaining network stability through adaptive traffic management.
3Reliability
If Explicit Congestion Notification is used to manage congestion, then packet loss is reduced through marking, but response time is delayed and performance suffers
Solution Approach 1:
Congestion prevention actions are taken preliminarily by isolating incast traffic into dedicated virtual routing fields before congestion occurs. This preliminary segmentation prevents congestion buildup, eliminating the need for reactive ECN marking and reducing response delays.
Solution Approach 2:
The patent converts the potentially harmful incast traffic pattern into a beneficial isolated flow within dedicated VRFs. By transforming the incast problem into a controlled virtual routing scenario, the system achieves both high throughput and low latency without relying on ECN's gradual response mechanism.
Data Source
AI summary
A switch including a plurality of ports; a management processor; and a switch core configured to receive a packet for transmission including a destination local ID (‘DLID’) and a virtual routing field (‘VRF) augmenting the routing of the packet on a route to the DLID according to a particular routing algorithm in dependence upon the VRF.