Cache-aware scheduling – The Linux patch aiming for +360% performance on MySQL

Source Domain: korben.info

The
Linux kernel scheduler [FR]
has just received a proposed update that delivers some pretty spectacular performance gains on certain workloads. Hygon, the Chinese chip maker that manufactures x86 processors under license from AMD’s Zen architecture, has sent a series of patches to extend cache-aware scheduling, and the announced numbers climb as high as 360% improvement in transactions per second on MySQL.

To understand what’s going on, we need to go back to basic
cache-aware scheduling
, the famous CAS, designed by Intel engineers (Tim Chen, Chen Yu, and Peter Zijlstra) and just merged into Linux 7.2. On a modern CPU with multiple last-level caches, the famous LLC, the scheduler tries to group tasks that share data onto the same cache domain. The result: fewer cache misses, less data bouncing between caches, and therefore better performance without touching the hardware – just by placing tasks more intelligently.

The catch is that this basic CAS only thinks in terms of a single LLC. As long as your workload fits within one cache domain, great. But the moment the workload exceeds what a single shared cache can hold, the scheduler has no way to group tasks at the level above: they spread across cores that no longer share the same cache, and all locality goes out the window. That’s a problem for Hygon, whose recent chips aren’t a single monolithic block but an assembly of chiplets (the C86-7490 brings together four of them), with multiple shared caches scattered across the die.

Hence the idea behind these patches, which enable hierarchical grouping and offer the ability to expand or contract dynamically depending on workload size and the machine’s topology.

Hygon is therefore claiming up to +49% on
Hackbench
, +20% on
Schbench
(no, not the rapper), and that famous +360% on MySQL!! That’s insane!

Now, before you sell your PC to grab a Hygon-powered one, hold on! These numbers are “up to” figures, measured on multi-domain topologies – meaning typically large servers…

Source