{"id":1934,"date":"2026-01-28T09:07:17","date_gmt":"2026-01-28T09:07:17","guid":{"rendered":"https:\/\/www.hammerdb.com\/blog\/?p=1934"},"modified":"2026-01-28T14:50:52","modified_gmt":"2026-01-28T14:50:52","slug":"scaling-databases-in-the-chiplet-era-lessons-from-mariadb-performance-engineering","status":"publish","type":"post","link":"https:\/\/www.hammerdb.com\/blog\/uncategorized\/scaling-databases-in-the-chiplet-era-lessons-from-mariadb-performance-engineering\/","title":{"rendered":"Scaling Databases in the Chiplet Era: Lessons from MariaDB Performance Engineering"},"content":{"rendered":"

I\u2019ll be presenting at MariaDB Day 2026 in Brussels on HammerDB performance tuning and benchmarking methodology that delivers real-world architectural impact,\u00a0 How MariaDB Achieved 2.5\u00d7 OLTP Throughput in Enterprise Server 11.8<\/a>. This blog post provides some of the insights and background that went into this work.<\/p>\n

\"\"<\/a><\/p>\n

In 2025 HammerDB began working with MariaDB engineering to analyse OLTP scalability on modern hardware platforms and to identify architectural bottlenecks that limit throughput at scale. The goal was not to tune a benchmark, but to use accurate, comparable and repeatable workloads to\u00a0 identify fundamental serialization points in the database engine. You can read the blog post here on the results of that work:\u00a0 Performance Engineered: MariaDB Enterprise Server 11.8 Accelerates OLTP Workloads by 2.5x<\/a><\/p>\n

To deliver improvements as quickly as possible, these changes were introduced in MariaDB Enterprise Server 11.8.3-1<\/strong>, with updates further progressing into the community version from MariaDB 12.2.X<\/strong> onwards.<\/p>\n

Note: It has been highlighted that some recent public benchmarks<\/a> of MariaDB evaluated earlier MariaDB 11.8 Community builds that did not yet include these scalability improvements. To observe the latest performance engineering benefits you should use the latest version of MariaDB Enterprise Server 11.8.3-1 or above<\/strong> or MariaDB Community Server 12.2.1<\/strong> (latest version at the time of publishing of this blog).”<\/p>\n

\n

1. Why Modern Hardware Changes Database Scaling<\/h2>\n

You may notice that the subtitle explicitly says MariaDB 11.8 delivers up to 2.5x higher OLTP throughput on Dell PowerEdge R7715 servers powered by AMD EPYC\u2122 processors<\/strong>.\u00a0 Database performance can never be viewed in isolation from the hardware. \u00a0If you have a familiarity with Moore\u2019s Law you will know that CPU transistor density historically doubled approximately every two years, leading to faster, cheaper, and more efficient computing power.\u00a0 However, in the chiplet era, performance scaling is no longer driven purely by transistor density\u2014system architecture, topology, and interconnect design now dominate real-world database scalability.<\/p>\n

Amdahl\u2019s Law, USL, and CPU Topology<\/h3>\n

The classic underlying scaling model applied by performance engineering is Amdahl\u2019s Law, which states that overall speedup is limited by the serial fraction of a workload. Even small serial sections cap scalability as core counts rise. The Universal Scalability Law extends this by modelling contention and coherency overheads, explaining why performance can plateau or even regress at high concurrency.<\/p>\n

In the chiplet era, CPU topology adds another dimension: NUMA boundaries and interconnect latency now create intra-CPU scaling knees that classical models did not explicitly predict. Modern database performance engineering therefore requires topology-aware analysis, not just thread-level scaling curves. For this reason we partnered with StorageReview<\/a> to access the very latest Dell system to evaluate MariaDB performance.<\/p>\n

The server platform used was a Dell R7715 equipped with:<\/p>\n