Artifact for the ASPLOS '26 paper — Performance Predictability in Heterogeneous Memory.

   ____    _    __  __ ____  
  / ___|  / \  |  \/  |  _ \ 
 | |     / _ \ | |\/| | |_) |
 | |___ / ___ \| |  | |  __/ 
  \____/_/   \_\_|  |_|_|    

CAMP: Causal Analytical Memory Prediction CAMP is a principled framework for accurately predicting application slowdown on heterogeneous memory systems combining DRAM and CXL. Using at most 12 hardware performance counters, it decomposes slowdown into three orthogonal components (demand reads, cache/prefetching, and stores). It can predict CXL/NUMA slowdown with requiring only a DRAM baseline run for latency-bound workloads. It also provides a closed-form model for weighted DRAM–CXL interleaving.

This repository contains the artifact code for reproducing the experiments in the paper. It is organized into two directories corresponding to the two CAMP models.

CAMP-Private/
├── prof/                # CXL slowdown prediction (DRAM run → predict CXL-induced stalls)
│   ├── <suite>/         # One directory per workload suite
│   ├── microbenchmarks/ # Calibration prediction models
│   └── proc/            # Data processing & model fitting (update_data.py, param.py, prof.py)
└── interleave/          # Weighted interleaving prediction (DRAM:CXL ratio sweep)
    ├── <suite>/         # One directory per workload suite
    ├── proc/            # Data processing & prediction (update_data.py, prof.py, pred.py)
    └── kernel-patch/    # Linux kernel patch for dynamic interleave weight control

• prof/ — Implements the CXL slowdown prediction model. Runs each workload in DRAM-only mode (local node 0) — plus a CXL run for bandwidth-bound workloads — and uses the captured PMU signals to analytically decompose and predict CXL-induced slowdown. Microbenchmarks in microbenchmarks/ calibrate prediction model coefficients via param.py.

• interleave/ — Implements the weighted interleaving prediction model. Sweeps DRAM:CXL interleaving ratios (plus all-DRAM L100 and all-CXL L0 endpoints) by writing weights to /sys/devices/system/node/node{0,1}/access0/il_weight via the included kernel patch, then provides a closed-form model for performance at any interleaving ratio.

• Linux with CXL/NUMA support and Intel PMU
• perf, numactl, vmtouch, libgfortran5, libxmu6
• sudo access (required for NUMA configuration, CPU frequency control, and perf events)
• Per-suite prerequisites listed in the Workload Suites section

Install common dependencies from either experiment directory:

./setup.sh

For interleave/ experiments, apply the kernel patch once to enable the il_weight sysfs knobs:

cd interleave/kernel-patch
patch -p1 < interleave.patch   # apply to your kernel source, then rebuild

1. Run a workload suite (e.g., cpu2017):

   cd prof/cpu2017
   sudo ./run.sh w.txt          # run all workloads
   sudo ./run.sh w.txt 1        # run only workload #1

2. Run microbenchmarks to calibrate the model:

   cd prof/microbenchmarks
   sudo ./run.sh w.txt

3. Calibrate model coefficients:

   cd prof/proc
   python3 update_data.py       # parse perf output → csv/
   python3 param.py             # fit slowdown model → params.txt

4. Analyze slowdown decomposition:

   python3 prof.py              # generate plots/ with per-component breakdown

1. Run a workload suite with the interleaving sweep:

   cd interleave/cpu2017
   sudo ./run.sh w.txt          # run all workloads (default: 9 interleaving ratios + L100 + L0)
   sudo ./run.sh w.txt 1        # run only workload #1

   Override the sweep resolution (default INTERLEAVE_PARTITIONS=10 yields 9 ratio points):

   sudo INTERLEAVE_PARTITIONS=5 ./run.sh w.txt 1   # 4 weight points

2. Post-process results:

   cd interleave/proc
   python3 update_data.py       # parse perf output → csv/all_data.csv
   python3 prof.py              # measured slowdown breakdown → plots/
   python3 pred.py              # predicted vs. measured + best-ratio report → plots/

All suites share the same run.sh interface. See interleave/README.md for full per-suite setup instructions.

Raw results land in rst/<workload>/ with the following extensions:

Processed outputs:

• csv/all_data.csv (interleave) or csv/mLOCAL.csv + csv/mNUMA.csv (prof)
• plots/<suite>__<workload>.png — slowdown breakdown
• plots/<suite>__<workload>__sd_*_pred.png — predicted vs. measured per component

@inproceedings{liu2026camp,
  title     = {Performance Predictability in Heterogeneous Memory},
  author    = {Liu, Jinshu and Xu, Hanchen and Berger, Daniel S. and Aguilera, Marcos K. and Li, Huaicheng},
  booktitle = {Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems},
  series    = {ASPLOS '26},
  year      = {2026},
  doi       = {10.1145/3779212.3790201}
}

This project is licensed under the MIT License — see the LICENSE file for details.