vSwitch LID Mapping for InfiniBand VM Migration

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In high-performance computing environments, particularly in InfiniBand networks, the challenges of live migration of virtual machines are hindered by complicated addressing and routing schemes, leading to connectivity issues and performance overhead during reconfiguration, especially with SR-IOV technologies.

Innovation Solution

The implementation of a vSwitch architecture with prepopulated and dynamic LID assignment mechanisms, allowing for efficient and scalable reconfiguration of virtual machines without the need for extensive network recalculations, thereby facilitating transparent live migration and reducing downtime.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional network addressing and routing schemes are used in InfiniBand networks with SR-IOV technology, then virtual machine migration is possible, but network recalculations become complicated and performance overhead increases

Engineering Contradiction:
Improvevirtual machine migration capabilityVSAvoidnetwork recalculation complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent pre-populates LID to HCA port mappings in the vSwitch before virtual machine migration occurs. This preliminary action stores the mapping information in advance, eliminating the need for complex network recalculations during migration. The subnet manager pre-establishes the relationship between LIDs and HCA ports, so when VMs migrate, the routing information is already available and no recalculation is needed.

Inventive Principle:
Principle #10Preliminary action

2Ease of operation

If dynamic LID assignment is implemented in vSwitch architecture, then live migration of virtual machines is facilitated, but network reconfiguration overhead increases

Engineering Contradiction:
Improvelive migration facilitationVSAvoidnetwork reconfiguration time
Core Design Contradiction:
Ease of operationVSLoss of time

Solution Approach 1:

The patent creates and maintains a copy of the LID to HCA port mapping table in the vSwitch. This copy allows the vSwitch to independently handle LID assignments and mappings without requiring real-time communication with the subnet manager during migration events. The copied mapping information is readily available locally, enabling fast LID reassignment when VMs migrate between HCAs, thus reducing network reconfiguration time.

Inventive Principle:
Principle #26Copying

3Measurement precision

If controlled re-cabling mode is activated for switch ports, then link testing can be performed, but network connectivity may be disrupted

Engineering Contradiction:
Improvelink testing capabilityVSAvoidnetwork connectivity stability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent implements a dynamic state machine for switch ports that can transition between different operational states including normal operation, controlled re-cabling mode, and link testing mode. When controlled re-cabling mode is activated, the port state changes to allow LID reassignment and link testing. The system dynamically adjusts the port behavior based on the current mode, enabling link testing while managing connectivity disruption through state-controlled transitions rather than permanent disconnection.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentUS10841195B2System and method for controlled re-cabling and link testing for switches and switch ports in a high performance computing network
Publication Date: 2020.11.17 ORACLE INT CORP
  • US10841195B2 patent drawing
  • US10841195B2 patent drawing
  • US10841195B2 patent drawing

AI summary

Systems and methods can provide a switch mode that is similar to “boot in progress”, but that is admin (e.g., administrator) controllable. This mode can ensure that an admin can perform re-cabling and cable test operations for any single switch in an operational subnet (e.g., an InfiniBand subnet) without causing unnecessary subnet re-routing and re-configuration operations.