Explore the design space of page management for multi-tiered memory systems in this 15-minute conference talk from USENIX ATC '21. Delve into the challenges posed by tiered memory systems comprising DRAM and SCM, and learn why current operating system page management schemes fall short in optimizing memory usage. Examine an in-depth analysis of Linux's page management extending NUMA to support systems with both DRAM and Intel's DCPMM. Discover the critical factors affecting performance in multi-tiered memory systems, including access locality and access tier. Investigate the proposed AutoTiering techniques, designed to effectively utilize multi-tiered memory systems. Gain insights into experimental results demonstrating significant performance improvements for various workloads compared to the stock Linux kernel, unlocking the full potential of multi-tiered memory hierarchies.
Exploring the Design Space of Page Management for Multi-Tiered Memory Systems
Overview
Syllabus
Intro
Large Memory Systems
Multi-tiered Memory Systems
Need for Page Placement in Multi-Tiered Memory
Need for Page Reclamation in Multi-Tiered Memory
Exploiting Access Tier and Locality
Demotion of Least Accessed Page
Estimating Least Accessed Page
Hiding Latency of Page Demotion
Experimental Environments
Performance Evaluation
Effectiveness of AutoTiering-CPM
Effectiveness of LAP classification
Conclusion
Taught by
USENIX
Related Courses
-
Johnny Cache - The End of DRAM Cache Conflicts in Tiered Main Memory Systems
-
Revisiting Secondary Indexing in LSM-based Storage Systems with Persistent Memory
-
Scalable and Effective Page-table and TLB Management on NUMA Systems
-
Efficient Memory Overcommitment for I/O Passthrough Enabled VMs via Fine-grained Page Meta-data Management
-
LLFree - Scalable and Optionally-Persistent Page-Frame Allocation
