Improving Performance of Flash Based Key-Value Stores Using Storage Class Memory as a Volatile Memory Extension

Explore a practical solution for improving performance and reducing costs of key-value stores running on DRAM-constrained servers using Storage Class Memories (SCM). Dive into the use of Intel® Optane™ PMem 100 Series SCMs in AppDirect mode to extend RocksDB's available memory, one of Facebook's largest key-value stores. Learn about the design of a hybrid cache in RocksDB that harnesses both DRAM and SCM hierarchically, and examine performance characterizations for three major RocksDB use cases at Facebook: WhatsApp, Tectonic Metadata, and Laser. Discover how this approach achieves up to 80% improvement in throughput and 20% improvement in P95 latency over existing small DRAM single-socket platforms, while maintaining significant cost improvements over large DRAM dual-socket platforms. Gain insights from this pioneering study of the DCPMM platform in a commercial data center environment, covering topics such as data center memory requirements, high-performance storage servers, SCM characteristics, RocksDB architecture, workload analysis, hybrid cache design, evaluation methodologies, and the limitations of SCM deployments.

Intro
Data Center Memory Requirements
High-Performance Storage Servers at Facebook
Storage Class Memory (SCM) Characteristics
Intel Optane DC Persistent Memory (DCPMM)
RocksDB Architecture
Facebook RocksDB Workload Characteristics
Intel Optane DC Persistent Memory Evaluations (WhatsApp)
Facebook RocksDB extended with SCM
Hybrid Cache Design and Components
DRAM First Admission Policy
Evaluation Setup and Implementations
Workload Creation
DRAM First Policy Throughput all workloads
DRAM First Policy Latency improvement
Deep Dive Analysis
Cost and Performance
Limitations of SCM Deployments
Conclusion
Acknowledgement