Non-Uniform Memory Architecture - A Nearly Unfathomable Morass of Arcana

Dive into a comprehensive conference talk exploring the intricacies of Non-Uniform Memory Architecture (NUMA) systems. Learn about the complexities and challenges of NUMA, including its impact on program performance, measurement difficulties, and inadequacies of standard profiling tools. Gain insights into identifying and troubleshooting NUMA-related issues, and discover solutions for problems such as poor scaling, large overhead, and low memory performance. Explore the history, performance implications, and debugging techniques for NUMA systems, with a focus on practical experiences and lessons learned from working with these architectures. Equip yourself with valuable knowledge to save time and effort when working on NUMA systems in enterprise computing environments.

Intro
Short Version
Long Version
What is NUMA
Intel Skylake
Uniform Memory Architecture
NonUniform Memory Architecture
Skylake
History
Multisocket systems
Why NUMA
Performance Implications
Asymmetry
Measurements
Memory Interface
Cross Node
Conclusions
Memory Latency
Accessing
Proximity
Interleaved
Debugging