Memory Tiering With CXL Access Logs to Reduce NUMA Scan Latency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing memory tiering systems face challenges with high latency due to non-uniform memory access (NUMA) balancing scans, which cause CPU resource wastage, page faults, and unpredictable access sequences, leading to suboptimal system performance.

Innovation Solution

The proposed systems and methods utilize a compute express link (CXL) device capability hint to avoid NUMA balancing scans by using access logs to determine page promotions and demotions between memory tiers, optimizing memory allocation based on access counters and data structures.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If NUMA balancing scans are used for memory tiering, then memory allocation can be optimized, but CPU resource wastage and latency increase

Engineering Contradiction:
Improvememory allocation efficiencyVSAvoidlatency
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The system performs preliminary actions by maintaining access logs that record memory access patterns before NUMA balancing scans are needed. When a page fault occurs, the system can quickly determine promotion/demotion decisions based on pre-collected access log data, avoiding the need for lengthy scanning operations and reducing latency while still achieving optimal memory allocation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The invention extracts only the necessary information (access counters and access sequences) from the access logs when needed, rather than performing comprehensive scans of all memory structures. This selective extraction of relevant data minimizes CPU resource wastage while maintaining effective memory tiering decisions.

Inventive Principle:
Principle #2Taking out (Extraction)

2Reliability

If NUMA balancing scans are performed, then memory tiering can be maintained, but CPU resource wastage increases

Engineering Contradiction:
Improvememory tiering maintenanceVSAvoidCPU resource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The access logs automatically record memory access patterns in the background without requiring active CPU intervention for data collection. The logs self-update with each memory access, allowing the system to maintain reliable memory tiering information without continuous CPU resource consumption during scanning operations.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

Access log data is collected and prepared in advance, so when memory tiering decisions need to be made, the CPU only needs to read pre-processed information rather than performing resource-intensive scans. This preliminary data preparation significantly reduces CPU resource wastage while maintaining tiering reliability.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If traditional memory tiering methods are used, then system performance can be maintained, but access sequencing accuracy decreases

Engineering Contradiction:
Improvesystem performanceVSAvoidaccess sequencing accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system implements feedback mechanisms by continuously monitoring and recording access sequences in the logs. This feedback information about actual access patterns allows the system to make more accurate promotion and demotion decisions, improving both access sequencing accuracy and overall system performance through data-driven memory management.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4645099A1Systems and methods for reducing latency in memory tiering
Publication Date: 2025.11.05 SAMSUNG ELECTRONICS CO LTD
  • EP4645099A1 patent drawingFigure 1
  • EP4645099A1 patent drawingFigure 2
  • EP4645099A1 patent drawingFigure 3

AI summary

Provided are systems (300), methods, and apparatuses for systems (300) and methods for reducing latency of memory tiering devices. In one or more examples, the systems (300), devices, and methods include determining a number of pages (335) in a first memory tier (330) satisfies a threshold; based on the number of pages (335) in the first memory tier (330) satisfying the threshold, obtaining from an access log (390) an access counter (394) and a first physical address (392) associated with a page (355) of a second memory tier (360); translating the first physical address (392) to a second physical address (380) associated with a host; and modifying, based on the access counter (394), a counter field of a first data structure (385).