Explore a 24-minute conference talk from USENIX NSDI '17 that introduces Infiniswap, an innovative remote memory paging system designed for RDMA networks. Dive into the challenges faced by memory-intensive applications and learn how Infiniswap efficiently disaggregates memory across a cluster. Discover the system's architecture, including slab-based management, decentralized placement, and eviction strategies. Understand how Infiniswap improves application performance and cluster memory utilization without modifying user applications or the OS. Examine the implementation details and evaluation results showcasing significant throughput improvements and latency reductions for various workloads on unmodified applications like VoltDB, Memcached, PowerGraph, GraphX, and Apache Spark. Gain insights into the system's scalability, limitations, and potential future developments in efficient memory disaggregation.
Overview
Syllabus
Intro
Memory-intensive applications
Performance degradation
Memory underutilization
Disaggregate free memory
What are the challenges?
System Overview
How to meet the design objectives?
Management unit: memory page?
Management unit: memory slab!
Which remote machine should be selected?
Slab eviction
Which slab should be evicted?
Power of multiple choices
Implementation
What are we expecting from Infiniswap?
Application performance
Cluster memory utilization
Limitations and future work
Conclusion
Data transmission & remote transparency
Evaluation
Taught by
USENIX
