Zhiqi Wang [王之祺]

Timeseries Management Systems

A thorough research on the main design decisions of the timeseries management systems, including the data model, memory data management, and persistent data management.

• Data model: To solve the data redundancy issue of the timeseries data from the same data source, we propose a unified data model for both tags and data samples of timeseries, with a novel compression mechanism and a two-level indexing design.
• Memory data management: To mitigate the memory overhead and maintain more timeseries with limited memory, we design a flexible inverted index that can dynamically adapt its structure to the memory pressure.
• Persistent data management: To achieve high insertion throughput of big timeseries data, we design a dynamic time-partitioned LSM-tree with high insertion throughput, decent space efficiency, and efficient out-of-order data handling.

LSM-Tree-Based Key-Value Stores with Hybrid Cloud Storage

LSM-tree-based key-value stores are widely used as the storage engines of big data systems. As the data volume scales up, it is a natural trend to deploy the system on the cloud. However, the existing LSM-tree designs can not adapt to cloud storage because of the huge performance gap. We design MirrorKV with a balanced read/write performance which separates keys and values into two mirrored LSM-trees for better data locality and read performance, and designs different compaction mechanisms for fast and slow storage to improve write performance.

A Monolithic Software/Hardware Co-Design Key-Value File System

To mitigate the metadata manipulation overhead and I/O amplification of the traditional file systems designed for block storage, we implement a file system with a key-value interface, which offloads the data management to our computational storage platform.

• Host-side key-value filesystem: It translates the file semantics (inode and page contents) to key-value commands correspondingly.
• Host storage communication: We customize the Linux NVMe driver to bypass the Linux block layer and transmit the key-value commands.
• Storage-side design: We carefully design the flash translation layer (FTL) to handle the received key-value commands and manage the physical area of the SSD.