How Multi‑Core CPUs and NVM Are Redefining Database Design

Modern Processors and New Storage Development

Since around 2005 CPU manufacturers shifted from increasing clock frequency to multi‑core designs because of power and manufacturing limits. Modern servers now commonly host dozens of cores, giving rise to the "many‑core" (or "crowd‑core") concept.

With many cores the memory‑wall problem becomes severe: a memory access that once cost one CPU cycle now costs hundreds, making memory latency a dominant performance bottleneck.

Non‑volatile memory (NVM) such as Intel’s 3D XPoint combines the persistence of storage with near‑DRAM latency. Its key traits are non‑volatility, low latency, high capacity, and asymmetric read/write performance.

Four main NVM technologies exist, with phase‑change memory (PCM) being the most mature. PCM switches between crystalline (low‑resistance) and amorphous (high‑resistance) states to represent bits.

Compared with flash, PCM offers two orders of magnitude lower read/write latency and two orders of higher endurance, while keeping comparable capacity. Its density is 2‑4× that of DRAM and its idle power is only 1 % of DRAM, making it attractive for data‑center workloads.

Database Systems on Modern Processors

Traditional DBMS designs were created for single‑core, disk‑I/O‑bound environments of the 1970s. In a multi‑core world, coarse‑grained locks, large critical sections, and disk‑oriented algorithms cause severe scalability problems.

Empirical studies (e.g., Carnegie‑Mellon 2010) show that many open‑source databases (MySQL, PostgreSQL) fail to scale on many‑core hardware, with most execution time spent in cache‑unfriendly code paths, lock management, and log handling.

Key optimizations include:

Column‑store techniques for OLAP workloads to improve cache utilization.

Cache‑friendly data structures (e.g., separating frequently accessed transaction visibility fields into a compact PGXACT struct).

Vectorized execution engines that process batches of tuples to reduce function‑call overhead and improve locality.

Lock‑striping and inheritance locks to shrink critical sections and reduce contention.

Database Systems for New Storage

NVM changes the storage hierarchy, prompting three integration strategies:

Use NVM as a fast disk replacement—no DBMS changes required.

Employ NVM as a log device—benefits from low‑latency writes while keeping traditional data buffers.

Fully redesign the DBMS to place data directly in NVM, eliminating separate log buffers.

CMU’s VLDB 2017 paper introduced write‑behind logging (WBL) , which writes dirty pages directly to NVM on commit and defers log persistence, achieving near‑instant recovery and ~30 % performance gains on NVM compared with SSD.

WBL records only a commit time interval (Cp, Cd) instead of full after‑image data, allowing the system to skip redo during recovery because committed data is already durable in NVM.

Summary

Application demand, industry data, and hardware advances drive DBMS evolution.

In the multi‑core, memory‑computing era, designers must prioritize scalability and data‑locality to overcome the memory‑wall.

NVM may soon allow DBMS architectures to abandon traditional I/O concerns and focus on extensibility.