How to Build a NVM DBMS (J. Arulraj, SIGMOD 2017)

[benclmnt]

Link to paper

Main Contribution

This paper summarizes research done on all layers needed to build a DBMS on NVM + points out areas where current memory-oriented / disk-oriented approaches need to be re-examined.

Summary

Context: CMU's Peloton database

NVM-properties

• byte-addresability
• high write throughput
• Read/Write Asymmetry (we want less writes as it take longer + device wearing)

Access Interfaces

NVM can be accessed directly using an NVM-aware memory allocator or through a special filesystem.

• Allocator Interface
  • CLWB + SFENCE API to flush (and prevent reordering, respectively) to NVM.
  • non-volatile pointers: pointer to NVM location is mapped to the same virtual location after a restart
  • guarantees atomicity of memory allocations
• Filesystem interface: Microsoft and Linux added direct access storage (DAX) support

Storage Manager

• Access methods: Bw-tree, hash tables
• Logging & Recovery: Write-Behind Logging (WBL), only use NVM to store log
• Data Placement: DRAM-NVM (DRAM to store hot tuples), how to evict cold data to disk?

Execution Engine

• Plan Executor: redesign algorithms to reduce writes (but incur more reads), i.e. segment sort, segmented Grace hash join
• Query Optimizer: differentiate reads and writes + take into account the convergence of sequential and random accesses, when to materialize?
• SQL Extension: allow user to control data placement e.g. in Oracle / SAP HANA