STEP Prefetcher

This repository contains the artifact used to evaluate STEP, the spatial footprint prefetcher proposed in:

STEP: Spatial Footprint Prefetcher with Multi-Point Temporal Triggers Yuanji Ye, Oliver Lenke, Thomas Wild, Andreas Herkersdorf

10.1109/ISCA66397.2026.00095

The codebase is built on top of ChampSim. This submission tree is a cleaned version of the development repository:

• traces are not included
• generated logs, JSON outputs, figures, and CSV results are not included
• run/build entry points are restricted to the experiment sets used in the paper

Repository Layout

• ChampSim/ ChampSim source tree and prefetcher implementations.
• scripts/make/ Build scripts for the paper experiments.
• scripts/experiments/ Run scripts and plotting scripts for the paper experiments.
• scripts/results_analysis/ Metric collection and shared plotting utilities.
• scripts/workloads.py Single-core workloads and heterogeneous multi-core mixes.
• traces/ Place downloaded trace files here.
• log/, json/, figures/ Created/used for generated outputs.

Tested Environment

The artifact was last validated on the following machine:

• OS: Ubuntu 24.04.4 LTS
• Kernel: Linux 6.17.0-1012-oem
• Compiler: g++ 13.3.0
• Python: 3.12.3
• Git: 2.43.0
• Python packages:
  • matplotlib 3.6.3
  • numpy 1.26.4
  • pandas 2.1.4
  • seaborn 0.13.2

Hardware Dependencies

The full campaign is compute- and storage-intensive. The evaluation machine used for the paper had:

• CPU: 2 x Intel Xeon Platinum 8462Y+
• Logical CPUs: 128
• Memory: 503 GiB RAM

Smaller machines can still run the scripts, but you should reduce --max-jobs substantially and expect much longer runtimes.

Software Dependencies

Install the standard build and scripting dependencies first.

sudo apt-get update
sudo apt-get install -y build-essential cmake git python3 python3-pip xz-utils
python3 -m pip install --user matplotlib numpy pandas seaborn

ChampSim uses vcpkg for its packaged dependencies:

cd ChampSim
git clone https://github.com/microsoft/vcpkg.git
./vcpkg/bootstrap-vcpkg.sh
./vcpkg/vcpkg install

Data Dependencies

The paper experiments use the workload lists defined in scripts/workloads.py.

• Single-core workload set:
  • 39 SPEC CPU2006 traces
  • 39 SPEC CPU2017 traces
  • 52 CloudSuite traces
  • Total: 130 traces
• Heterogeneous multi-core workload sets:
  • 50 random 2-core mixes
  • 50 random 4-core mixes
  • 50 random 8-core mixes

Trace sources:

• SPEC CPU2006 / SPEC CPU2017 ChampSim traces: DPC-3 trace archive
• CloudSuite traces: CRC-2 trace archive (currently redirects to a Dropbox folder)

These are the same public ChampSim trace sources referenced by the Pythia artifact and the upstream ChampSim documentation. The archives are maintained outside this repository and their availability is controlled by their respective hosts. Keeping a local mirror is recommended.

Create the trace directory and download the files listed in scripts/workloads.py:

mkdir -p traces
# Download the required .champsimtrace.xz files from the archives above.
# Keep the original archive filenames and place the files directly in traces/.

The experiment scripts resolve traces as traces/<filename>. Before launching the full campaign, check that all 130 expected single-core traces are present:

python3 - <<'PY'
from pathlib import Path
from scripts.workloads import workloads_all

trace_dir = Path("traces")
missing = [name for name, *_ in workloads_all if not (trace_dir / name).is_file()]
print(f"Expected: {len(workloads_all)}; found: {len(workloads_all) - len(missing)}; missing: {len(missing)}")
if missing:
    print("\n".join(missing))
    raise SystemExit(1)
PY

Trace files can be large and may have separate distribution terms. They are ignored by Git and are not covered by this repository's MIT license.

Setup

From the repository root:

cd /path/to/STEP

Build the binaries required by the paper:

python3 scripts/make/make_all.py

This compiles the binary sets used by:

• single-core L2 performance
• single-core L1-level experiments
• multi-core homogeneous and heterogeneous experiments
• multi-level prefetching
• ablation study
• parameter sweep
• system sensitivity
• storage sensitivity
• limited-way L2 prefetching

Experiment Coverage

• Single-core L2: baseline no, SMS, DSPatch, SPP-PPF, PMP, vBerti, Gaze, eBingo, STEP
• Multi-core: baseline no, DSPatch, Gaze, eBingo, STEP
• L1-level: baseline no, Gaze, eBingo, STEP
• Multi-level: all L1/L2 combinations of Gaze, STEP, eBingo, vBerti except vBerti+vBerti, plus IPCP
• Ablation: step_full_ft256_at128_pt8, step_disable_first_ft256_at128_pt8, step_disable_second_ft256_at128_pt8, step_disable_third_ft256_at128_pt8
• Parameter sweep: STEP variants with
  • FT in {32, 64, 128, 256, 512, 1024} at AT=128, PHT=8
  • AT in {32, 64, 128, 256, 512} at FT=256, PHT=8
  • PHT in {4, 8, 16, 32, 64, 128} at FT=256, AT=128
• System sensitivity: baseline no, Gaze, vBerti, SPP-PPF, eBingo, STEP
• Storage sensitivity: storage variants of STEP, Gaze, vBerti, Bingo, eBingo, and IPCP
• Limited-way L2: baseline no, Gaze, eBingo, and STEP, with prefetches restricted to one L2 way

Running Experiments

All paper experiments can be launched together with the shared global queue:

python3 scripts/experiments/run_all_experiments.py --max-jobs 32

By default, completed runs are skipped. Use --no-skip-completed if you need to rerun everything.

You can also run each experiment independently.

Single-Core L2 Performance

python3 scripts/experiments/single_core_performance/run_single_core.py --max-jobs 32

L1-Level Experiment

python3 scripts/experiments/l1_level/run_l1.py --max-jobs 32

Multi-Core Performance

Homogeneous:

python3 scripts/experiments/multi_core_performance/run_multi_core.py --max-jobs 32

Heterogeneous:

python3 scripts/experiments/multi_core_performance/run_multi_core.py --heterogeneous --max-jobs 32

Multi-Level Prefetching

python3 scripts/experiments/multi_level_prefetching/run_multi_level_prefetching.py --max-jobs 32

Ablation Study

python3 scripts/experiments/ablation_experiment/run_ablation.py --max-jobs 32

Parameter Sweep

python3 scripts/experiments/parameter_experiment/run_parameter_sweep.py --max-jobs 32

System Sensitivity

python3 scripts/experiments/system_parameter_experiment/run_system_sensitivity.py --max-jobs 32

Storage Sensitivity

python3 scripts/experiments/storage_sensitivity/run_storage_sensitivity.py --max-jobs 32

Limited-Way L2 Prefetching

python3 scripts/experiments/limited_way_experiment/run_limited_way_experiment.py --max-jobs 32

Collecting Metrics

After the simulations finish, regenerate the standard CSV files with:

python3 scripts/results_analysis/collect_all_metrics.py

This produces the expected CSV outputs under scripts/results_analysis/results/.

If you only need one experiment, use scripts/results_analysis/collect_metrics.py directly with the appropriate --log-root and --output arguments.

Plotting

Each experiment directory contains its corresponding plot script. Typical commands:

Single-Core L2

python3 scripts/experiments/single_core_performance/plot_single_core_metrics.py

L1-Level

python3 scripts/experiments/l1_level/plot_l1_level_metrics.py

Multi-Core Scaling

python3 scripts/experiments/multi_core_performance/plot_multicore_scaling.py

Multi-Level Prefetching

python3 scripts/experiments/multi_level_prefetching/plot_multi_level_prefetching.py

Ablation

python3 scripts/experiments/ablation_experiment/plot_ablation.py

Parameter Sweep

python3 scripts/experiments/parameter_experiment/plot_parameter_sweep.py

System Sensitivity

python3 scripts/experiments/system_parameter_experiment/plot_l2_system_sensitivity.py

Storage Sensitivity

python3 scripts/experiments/storage_sensitivity/plot_storage_sensitivity_metrics.py

Limited-Way L2 Prefetching

python3 scripts/experiments/limited_way_experiment/plot_limited_way_experiment.py --reserved-ways 1

Generated figures are written under each experiment's figures/ directory unless an explicit output path is provided.

Notes

• Use the same warmup/simulation counts for build, run, collection, and plotting workflows. The default paper configuration is 50M warmup and 100M simulation instructions.
• The queue-based runners accept --max-jobs, --poll-seconds, and retry controls so you can adapt them to your machine.
• The Git tree intentionally excludes traces and generated outputs; those artifacts are regenerated locally.

License

STEP-specific source code and experiment scripts are released under the MIT License. The bundled ChampSim source tree and third-party prefetcher implementations retain their respective upstream copyright and license notices; see ChampSim/LICENSE.