NUMA Memory Access via Global Physical Address Synchronization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-Uniform Memory Access (NUMA) systems face increased access latency and reduced performance due to remote memory access, particularly when system software like operating systems or virtual machine monitors need to share codes and data across nodes, leading to complex page table management and remote memory access burdens.
Innovation Solution
Implementing a memory access method where each node in a NUMA system uses a shared global physical address for local memory, synchronizing operations across nodes to avoid remote memory access by using synchronization messages for data allocation, update, and release, thereby ensuring a continuous global physical address space and reducing the need for page table synchronization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If remote memory access is used in NUMA system, then memory sharing between nodes is enabled, but access latency increases and performance decreases
Solution Approach 1:
The patent segments the global physical address space into node-specific local address spaces. Each node maintains its own local memory with a local address space that maps to a portion of the global address space. This segmentation allows each node to access its local memory directly without remote access, while still enabling memory sharing through the global address space mapping mechanism.
Solution Approach 2:
The patent implements local quality by allowing each node to have its own local memory and local address space mapping. Instead of uniform remote access to a centralized memory, each node accesses its local memory through its own address space, optimizing access performance while maintaining global address space consistency through synchronization.
2Reliability
If page table migration and backup operations are implemented, then memory address consistency is maintained, but system complexity and operational overhead increase
Solution Approach 1:
The patent applies preliminary action by pre-establishing the mapping relationship between global physical address space and each node's local address space at system initialization. The mapping relationships are configured in advance, and synchronization mechanisms are pre-set up, so that when memory operations occur, the system can directly use the pre-configured mappings without complex runtime migration or backup operations.
3Stability of the object's composition
If centralized memory management is used, then address space continuity is maintained, but remote access operations increase and performance decreases
Solution Approach 1:
The patent introduces a new dimension by creating local address spaces at each node that map to the global address space. Instead of a single centralized memory dimension, the system uses multiple local memory dimensions (one per node), each with its own address space. This dimensional transformation allows the system to maintain global address space continuity while enabling local access performance through the additional local dimension.
Data Source
AI summary
A memory access method and device are provided for a Non-Uniform Memory Access (NUMA) system including nodes, each node configured to run a predetermined program and including a preset memory allocated for the predetermined program, where a same global physical address corresponds to each of the preset memories of each of the nodes, respectively, the method including: performing an operation on a first preset memory local to a first node of the nodes when the predetermined program is running on the first node; and synchronizing the operation performed on the first preset memory to each preset memory local to each of the nodes, respectively, where each of the nodes is configured to access its local preset memory using the same global physical address when running the predetermined program on the node.