Domain-Aware Data Migration in CC-NUMA Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current cache-coherent multi-processor systems lack domain awareness, leading to inefficient data movement due to the absence of affinity information being propagated downstream, resulting in suboptimal performance and increased latency in CC-NUMA systems.
Innovation Solution
A domain-aware data migration method is introduced, which generates and distributes ID maps (IDM) and system address maps (SAM) tables at boot-time to provide locality information to processing and memory elements, enabling intelligent data movement between compute, memory, and caches, thereby aligning hardware behavior with software affinity optimizations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data migration operations are performed without domain awareness and affinity information propagation, then system implementation is simpler, but data movement efficiency deteriorates and latency increases
Solution Approach 1:
The patent applies preliminary action by generating and distributing IDM and SAM tables at boot-time, before data migration operations commence. This pre-computation of locality information enables subsequent data migration to be efficient without real-time complexity, as the hardware already possesses the necessary affinity information when data movement operations begin.
Solution Approach 2:
The patent introduces intermediary data structures (IDM and SAM tables) that mediate between the software layer which knows affinity information and the hardware layer which executes data migration. These tables serve as intermediaries that translate software-aware affinity information into hardware-accessible locality information, enabling efficient data movement without requiring complex real-time communication between software and hardware.
2Speed
If locality information is propagated downstream to hardware elements, then data migration performance improves, but information propagation complexity increases
Solution Approach 1:
The patent applies preliminary action by generating and distributing IDM and SAM tables at boot-time, before data migration operations commence. This pre-computation of locality information enables subsequent data migration to be efficient without real-time complexity, as the hardware already possesses the necessary affinity information when data movement operations begin.
Solution Approach 2:
The patent segments the affinity information into two distinct data structures: IDM tables that map processing element IDs to locality domains, and SAM tables that map memory address ranges to locality domains. This segmentation allows hardware elements to access only the relevant portion of affinity information needed for their specific operations, reducing the complexity of information propagation while maintaining high data migration speed.
3Productivity
If affinity information is made available to software layers, then software optimization is improved, but hardware awareness remains lacking
Solution Approach 1:
The patent introduces intermediary data structures (IDM and SAM tables) that mediate between the software layer which knows affinity information and the hardware layer which executes data migration. These tables serve as intermediaries that translate software-aware affinity information into hardware-accessible locality information, enabling efficient data movement without requiring complex real-time communication between software and hardware.
Solution Approach 2:
The patent implements feedback by propagating locality information downstream from the software layer to the hardware layer. The IDM and SAM tables carry affinity information that was originally determined by software (based on UEFI/ACPI standards) back to the hardware elements, creating a feedback loop that ensures hardware behavior aligns with software optimization decisions.
Data Source
AI summary
Embodiments disclosed herein provide a domain aware data migration scheme between processing elements, memory, and various caches in a CC-NUMA system. The scheme creates domain awareness in data migration operations, such as Direct Cache Transfer (DCT) operation, stashing operation, and in the allocation of policies of snoop filters and private, shared, or inline caches. The scheme defines a hardware-software interface to communicate locality information (also referred herein as affinity information or proximity information) and subsequent hardware behavior for optimal data migration, thus overcoming traditional CC-NUMA limitations.


