NVMe RDMA Arbitration Mapping for Queue Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In NVMe-aware RDMA networks, the mismatch in arbitration policies between host RNIC, target RNIC, and NVMe-aware SSDs leads to inefficiencies in data packet transmission due to legacy RDMA features' inability to comprehend advanced NVMe features like Weighted Round Robin for queue arbitration.
Innovation Solution
The solution involves storing queue arbitration settings of RDMA queue pairs in an arbitration mapping table for both host and target RNICs, matching them with the NVMe-aware SSD's settings, allowing for alignment and efficient data packet transmission through virtual lanes using a vendor-specific arbitration format.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If NVMe-aware SSDs generate their own queue pairs and arbitration policies, then the SSD can independently manage its queues, but this leads to mismatch in arbitration policies between host RNIC, target RNIC, and SSD
Solution Approach 1:
The patent introduces an arbitration mapping table as an intermediary structure that stores and aligns arbitration policies between host RNIC, target RNIC, and NVMe-aware SSD. This mapping table acts as a mediator that ensures consistent queue pair arbitration across all three components, preventing policy mismatches while maintaining independent queue management capabilities.
Solution Approach 2:
The patent changes the arbitration policy parameters by storing weights and priorities of RDMA queue pairs in the arbitration mapping table. This allows the system to dynamically adjust and align arbitration parameters across different components, ensuring consistency while maintaining adaptability.
2Adaptability or versatility
If legacy RDMA features are used for queue arbitration, then compatibility is maintained, but advanced NVMe features like Weighted Round Robin cannot be comprehended
Solution Approach 1:
The arbitration mapping table serves as an intermediary that translates and aligns advanced NVMe arbitration policies (like Weighted Round Robin) with legacy RDMA queue pair structures. This allows the system to comprehend and implement advanced NVMe features while maintaining compatibility with existing RDMA infrastructure.
Solution Approach 2:
The patent creates a universal arbitration mechanism that can handle both legacy RDMA queue arbitration and advanced NVMe arbitration policies through the same mapping table structure. This multi-functional approach allows the system to support multiple arbitration schemes without increasing overall system complexity.
3Productivity
If queue pairs are not aligned between host RNIC and target RNIC, then each component can use its own arbitration policy, but data packet transmission efficiency decreases
Solution Approach 1:
The arbitration mapping table acts as an intermediary structure that aligns queue pairs between host RNIC and target RNIC. By storing and referencing consistent arbitration policies in this mapping table, the system ensures proper queue pair alignment without requiring complex coordination mechanisms between the two RNICs.
Solution Approach 2:
The patent performs preliminary alignment of queue pairs and arbitration policies by populating the arbitration mapping table before data transmission begins. This preliminary action ensures that queue pairs are properly aligned and arbitration policies are consistent, eliminating the need for complex runtime coordination and improving transmission efficiency.
Data Source
AI summary
The present disclosure describes a method and a system for sending data packets to improve Quality of Service in Non-Volatile Memory express (NVMe) aware Remote Direct Memory Access (RDMA) network, including receiving, by a host RNIC, RDMA command from a host initiator, wherein the RDMA command comprises data packets, arranging, by the host RNIC, the data packets based on weights and priorities of RDMA queue pairs, storing, by the host RNIC, the data packets in a host queue from host RDMA queue pairs based on the weights and priorities of the RDMA queue pairs, and sending, by the host RNIC, the data packets through host virtual lanes to a target RNIC.