LMS Predictor Tackles Memory Bandwidth Contention in Multicore Linux

Memory bandwidth contention is the silent killer of predictable performance in multicore systems. LMS-AR, a new Linux kernel module from researchers on GitHub, uses a least-mean-square adaptive filter to predict exactly how much bandwidth each core needs and allocates accordingly.

Master Core Runs the Predictor, Cores Stay Fast

The design sidesteps the usual chicken-and-egg problem: monitoring and regulation happen from a master core that is not a dedicated controller. That master core runs the LMS prediction algorithm - computationally cheap enough for real time - without stealing cycles from the regulated cores. Per-core bandwidth requirements are forecast using an adaptive filtering technique, then enforced through the kernel's existing resource control mechanisms.

SPEC CPU 2017 Benchmarks Show Clear Gains Over Memguard

Experiments with SPEC CPU 2017 benchmarks spread across multiple cores showed that LMS-AR significantly improved slowdown ratios compared to Memguard, a well-known memory bandwidth regulator. No exact percentages were disclosed, but the results are consistent enough to warrant a public release. The code is on GitHub at https://github.com/ss22ongithub/LMSAdaptiveRegulator.

What this means: you can now plug an adaptive predictor into production Linux kernels to tame memory interference without dedicating a hardware core to the job. LMS-AR turns memory from a noisy shared resource into something you can schedule like CPU time.

Source: LMS-AR: LMS Prediction-based Adaptive Regulator for Memory Bandwidth in Multicore Systems
Domain: arxiv.org