Overview
Explore a conference talk from OSDI '22 that introduces MemLiner, a runtime technique designed to enhance far-memory systems' performance. Learn how this innovative approach aligns memory accesses from applications and garbage collection to reduce local-memory working sets and improve remote-memory prefetching. Discover the challenges faced in modern datacenters regarding large memory footprints and resource utilization, and understand how MemLiner addresses these issues. Examine the implementation of MemLiner in OpenJDK's G1 and Shenandoah garbage collectors, and review the impressive performance improvements achieved in widely-deployed cloud systems. Gain insights into object classification, barriers, and the key design ideas behind this award-winning research that promises to revolutionize far-memory techniques in high-level language environments.
Syllabus
Intro
Memory Capacity Bottleneck in Datace
Far-Memory System
High-level Languages
Garbage Collection
Resource Competition
Ineffective Prefetching
Can we disable concurrent tracing?
Observations
Key Design Idea
Object Classification
Challenges in Classifying Objects
Barriers
Local Objects
Incoming Objects
Distant Objects
Results: Prefetching Effectiveness
Key Takeaways
Taught by
USENIX