S4FIFO Control-Plane Model (compact)

A compact, dependency-free binary export of the S4FIFO learned control-plane model: a 20-model LightGBM multiclass ensemble (140,400 trees total) that predicts which of 18 candidate S4FIFO cache-eviction configurations best fits a workload, given 73 cache-level features collected at runtime.

This is a re-serialization of the model served at hxia7/s4fifo-api (a 372MB joblib/scikit-learn ensemble) into a flat, self-describing binary format (s4fifo_model.s4m, 189MB) that a small (250 line), dependency-free C loader/evaluator can read directly - no Python, scikit-learn, or LightGBM runtime required. It is bit-for-bit equivalent to the original: validated to pick the identical class on every one of 14 real production-trace feature vectors tested against the live API.

Usage

Used by libCacheSim's s4fifo eviction algorithm's optional learned control plane:

./cachesim trace.oracleGeneral oracleGeneral s4fifo 100mb \
  -e "auto-tune=1,model-path=/path/to/s4fifo_model.s4m"

(built with -DENABLE_S4FIFO_LEARNED=ON)

Format

See libCacheSim/cache/eviction/S4FIFO_model_real.c's header comment in the libCacheSim repo for the exact binary layout, and scripts/s4fifo_export_model.py for the exporter that produced this file from the original joblib ensemble.

Provenance

  • Trained as part of "Learning-Augmented Heuristics: Simple, yet Smart, Robust and Interpretable Cache Eviction" (OSDI 2026).
  • Original model: hxia7/s4fifo-api
  • 18-class candidate configuration table and cost matrix unchanged from the original.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support