Why Memory Databases Outperform Disk‑Based Systems: Key Technologies Explained

Disk‑Based DBMS

Traditional DBMS store data on disk because early hardware limited memory; systems like Oracle and MySQL still use this architecture, which incurs high I/O latency.

Why Memory Is Now Viable

With cheap, large‑capacity RAM (hundreds of GB to TB), entire structured datasets can reside in memory, eliminating many disk‑I/O bottlenecks for typical business workloads.

Buffer‑Pool Management vs Direct Memory Access

In disk‑based systems a page is read into a buffer pool, addressed via Page ID + Offset, and later written back; even with a full buffer pool the address‑translation overhead remains.

Lock vs Latch

Disk DBMS keep locks in a separate lock table, while memory DBMS embed lock information in record headers; latch protects internal data structures, lock protects logical data.

Logging and Recovery

Both systems use Write‑Ahead Logging (WAL). Disk DBMS employ “Steal + No‑Force” to balance durability and performance; memory DBMS only need redo logs because undo is unnecessary when data never leaves memory.

Performance Overhead of Disk‑Based DBMS

A 2008 SIGMOD study showed that only ~7 % of CPU cycles handle actual business logic, while the rest is spent on buffer management, latching, locking, logging, and B‑tree processing.

Historical Development of Memory Databases

Three stages: 1984‑1994 (research on in‑memory techniques), 1994‑2005 (first commercial in‑memory products such as Dali and TimesTen), and post‑2005 (modern high‑performance systems).

Key Architectural Innovations

Direct address access eliminating buffer‑pool indirection.

Data partitioning and functional partitioning.

Lock‑free and cache‑conscious designs.

Coarse‑grained locking (now less common).

Compiled query execution to avoid iterator overhead.

Scalable high‑performance index construction with reduced logging.

Modern Hardware Context

Large RAM, many‑core CPUs, and multi‑socket servers shift bottlenecks from I/O to CPU and runtime overhead, prompting removal of traditional buffer pools and adoption of compiled execution.

Conclusion

The article compares disk‑based and memory‑based DBMS architectures, outlines the evolution of in‑memory databases, and highlights the techniques that enable today’s high‑performance memory database systems.