RDMA Memory Replication via Page Pool Negotiation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current RDMA-based memory replication methods in data centers face performance limitations due to high latency and resource consumption, particularly in scenarios involving multiple messages for data transfer and expansive memory allocations, as well as poor error recovery and support for active-active implementations.
Innovation Solution
A method and system for efficient memory replication using RDMA writes with a memory negotiation scheme, employing an RDMA page pool, reconnect synchronization, and error recovery mechanisms via posted, retry, and pending queues, which allows for persistent and low-latency data replication with improved resource utilization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional storage protocols with multiple messages are used for memory replication, then data transfer can be achieved, but latency increases and processor overhead increases
Solution Approach 1:
The patent segments the replication process into distinct phases: connection establishment, memory region negotiation, and data transfer. By dividing the protocol into these segments, the system can optimize each phase independently, using RDMA for the data transfer phase to minimize latency while maintaining reliability through proper connection management in earlier phases.
Solution Approach 2:
The patent performs preliminary actions by establishing connections and negotiating memory regions before actual data transfer begins. This preliminary setup includes allocating memory buffers and agreeing on transfer parameters, which allows the subsequent RDMA operations to proceed immediately without additional negotiation overhead, thus reducing overall latency.
2Reliability
If conventional storage protocols with multiple messages are used for memory replication, then data transfer can be achieved, but processor overhead increases
Solution Approach 1:
The patent introduces RDMA as an intermediary mechanism that enables direct memory-to-memory transfers between nodes without involving the CPU in the data movement path. The RDMA hardware handles the transfer operations, protocol management, and error handling, freeing the processor to perform other useful work while maintaining reliable data transfer.
Solution Approach 2:
The system implements self-service by enabling memory replication operations to execute autonomously through RDMA hardware and zero-copy mechanisms. The source and target nodes can perform replication operations independently without requiring continuous processor intervention, allowing the system to maintain reliability while minimizing active processor involvement during data transfer.
3Productivity
If expansive memory allocations are performed for replication, then data transfer capacity increases, but resource consumption increases
Solution Approach 1:
The patent implements dynamic memory allocation where memory regions are negotiated and allocated based on actual replication needs rather than pre-allocating fixed large buffers. The system can dynamically adjust the size and number of memory regions during operation, allocating only the necessary memory resources for current replication tasks and releasing them when no longer needed, thus maintaining high transfer capacity while optimizing resource consumption.
Data Source
AI summary
A mechanism is provided in a data processing system comprising at least one processor and at least one memory, the at least one memory comprising instructions that are executed by the at least one processor and configure the at least one processor to implement a replication protocol for replicating memory from an initiator to a target. The initiator requests one or more pages of memory at the target. Responsive to the initiator receiving a page advertisement from the target, the initiator updates a list of pages at the initiator. The list of pages is replicated at the target. The initiator performs a remote direct memory access (RDMA) write to the one or more pages of memory at the target. Responsive to successful completion of the RDMA write, the initiator updates the list of pages at the initiator. Upon completion of the RDMA write the list of pages is updated at the target.