RDMA Session Fail-Over Module for Adapter Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing RDMA technologies lack effective mechanisms for transparently recovering from adapter failures, leading to disruptions in data transfer and coherence issues across networked systems, particularly in scenarios involving multiple adapters and diverse topologies.
Innovation Solution
The implementation of a session fail-over module that manages RDMA connections by creating redundant Queue Pairs, Memory Windows, and Completion Queues across primary and secondary adapters, ensuring transparent fail-over and maintaining data coherence through mechanisms like the session fail-over header and follow-on RDMA-read operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If RDMA adapter failure recovery mechanisms are implemented, then system reliability is improved, but device complexity increases
Solution Approach 1:
The system performs preliminary actions by creating backup Queue Pairs and completing pending operations before adapter failure occurs. The RNIC saves completion information to system memory in advance, and the host software prepares recovery sequences beforehand, enabling transparent failover without application intervention.
Solution Approach 2:
The patent introduces an intermediary recovery mechanism where the RNIC and host software coordinate to manage adapter failures. The RNIC acts as an intermediary that captures completion information and passes it to the host, which then orchestrates the recovery process by recreating Queue Pairs and restoring data coherence.
2Ease of operation
If transparent fail-over mechanisms are implemented, then ease of operation is improved, but device complexity increases
Solution Approach 1:
The system implements self-service by enabling the RNIC and host software to automatically detect adapter failures and execute recovery procedures without application software intervention. The completion information is automatically saved, and the recovery process is autonomously managed through coordinated host-RNIC operations.
Solution Approach 2:
Transparent fail-over is achieved through preliminary actions where the system pre-saves completion information and pre-establishes backup Queue Pairs before failures occur. This allows the fail-over mechanism to execute seamlessly without requiring application software to be aware of or intervene in the recovery process.
3Reliability
If data coherence is maintained through follow-on RDMA-read operations, then data integrity is improved, but productivity decreases
Solution Approach 1:
The system uses feedback through follow-on RDMA-read operations to verify data coherence after adapter recovery. The RNIC sends read requests to confirm that data was properly flushed to remote memory, and only after successful verification does the system resume normal operations, ensuring integrity while minimizing performance impact.
Solution Approach 2:
The patent applies partial action by performing selective RDMA-read operations only on specific data regions that require coherence verification, rather than flushing entire memory spaces. This targeted approach maintains data integrity for critical regions while minimizing the performance overhead associated with comprehensive data verification.
Data Source
AI summary
An apparatus in an example comprises an RDMA (Remote Direct Memory Access) network adapter located at an interface of a requester node coherency domain that sends an RDMA read request to a responder node coherency domain to flush into the responder node coherency domain data previously sent by the RDMA network adapter.