CacheFlex Artifact

This artifact provides the code, fixed configurations, reference results, and workflows for Figure 2, Figure 7, Figure 8, Figure 9, and Table 4 (flash scaling). Kernel provenance and the result-comparison policy are documented in docs/CLAIMS.md.

Quick check from reference results

The no-simulation check needs only Python 3.11--3.12 and the packages in requirements.txt:

python3 -c 'import sys; sys.exit(0 if sys.version_info[:2] in {(3, 11), (3, 12)} else "ERROR: CacheFlex requires Python 3.11 or 3.12")'
python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
source setup_env.sh
bash scripts/render_reference.sh

It does not require the ARM toolchain, QEMU, a gem5 build, the ROI-marker object, or the smoke tests. The environment command may therefore report those full-run components as missing; they are not used by this check. On a typical host, rendering the supplied results takes only a few minutes.

The script verifies all shipped reference manifests and result cells, including the complete Figure 9 grid (the Table 4 companion pool is validated by the configuration and value checks in plot_table4_flash.py --check), and renders:

• fig2_real.pdf, fig2_real.png, and fig2_data.json
• fig7_real.pdf, fig7_real.png, and fig7_data.json
• fig8_real.pdf, fig8_real.png, and fig8_data.json
• fig9_real.pdf, fig9_real.png, and fig9_data.json
• table4_data.json, table4_flash.tex, and table4_flash.md (Table 4, rendered from the frozen reference and the self-contained companion pool in experiments/end2end/table4/)
• OUTPUT_MANIFEST.sha256, binding every rendered figure to its plot data

All outputs go to results/generated/reference_outputs/. The command never modifies results/reference/ or any supplied reference figure.

Run all paper experiments

Complete SETUP.md once for the full simulation path. It installs the required tools, builds gem5, and checks the environment.

From the repository root, run:

bash run_all.sh -j 4

-j N sets the maximum total number of gem5 simulations that may run in parallel. Use a value from 1 to 16; the default is -j 1, and the effective value is also capped at the host's reported CPU count.

Expect roughly 1–2 days of wall clock for the full run at -j 12–-j 16 on a recent server: the largest individual cells (vl4 LLaMA T=4096 FFN/attention) each take several hours, and a fully loaded run keeps every slot busy. Individual cells are capped at 24 h wall clock (RH_CELL_TIMEOUT seconds to override).

The command builds the benchmarks and then runs the paper experiments in this order:

1. Figure 2: capacity sensitivity
2. Figure 7: SIMD-width sensitivity
3. Figure 8: software alternatives
4. Figure 9: component-composed prefill layers

The four stages run 21, 42, 12, and 108 selected cells, respectively. The default path runs only the selected configurations in each best.json; it does not perform parameter sweeps. The frozen SPM issue configuration uses two load slots at VL=4/8 and one load slot at VL=16, with one store slot at every vector length.

The regenerated paper outputs are written under results/generated/paper_outputs/:

• fig2_real.pdf, fig2_real.png, and fig2_data.json
• fig7_real.pdf, fig7_real.png, and fig7_data.json
• fig8_real.pdf, fig8_real.png, and fig8_data.json
• fig9_real.pdf, fig9_real.png, and fig9_data.json
• OUTPUT_MANIFEST.sha256, binding every rendered figure to its plot data

After plotting, run_all.sh compares the fresh time, performance, energy, speedup, and normalized-ratio values with the supplied paper reference. The fixed limit is 5% relative error. Figure 9 metrics are gated one-sidedly: regressions beyond 5% fail, while improvements of any size pass and are printed as [IMPROVED], because the shipped kernels include disclosed post-submission optimizations — see docs/CLAIMS.md for the kernel provenance and comparison policy. Figure structure, workload and cell identities must match exactly, and the existing correctness checks remain strict. A failed comparison returns a nonzero status and keeps the fresh outputs for inspection.

To run an experiment's declared tuning grid, add --sweep to its run.sh. These optional grids are not part of the default paper workflow. On a fresh code-only extraction, use the root run_all.sh workflow so the required benchmark binaries are built first.

License

CacheFlex's own contributions (kernels/, scripts/, spm_tools/, experiments/, and the CacheFlex modifications) are released under the MIT License; see LICENSE. The SVE FP16 GEMM microkernels in kernels/ are derived from the Arm Compute Library (Arm Limited, MIT License); that upstream notice is retained in the affected source files and in THIRD_PARTY_NOTICES. The bundled gem5 simulator (gem5/) and its third-party components (gem5/ext/) remain under their original licenses; see gem5/LICENSE and the respective subdirectories.