Chameleon-ATC

Chameleon reclaims and restores virtual-machine memory using page-access sampling, hierarchical hotness tracking, pressure feedback, and batched EPT updates. This artifact builds on HyperAlloc, runs Linux 6.18 in both Guest and Host, and uses Hermit over physical RDMA for remote memory. Guests are managed directly with QEMU.

The artifact evaluates two main claims:

1. Single VM (Figures 7 and 8): Chameleon provides a better trade-off between memory reclamation and application slowdown than the supplied baselines.
2. Multiple VMs (Figure 9): Chameleon achieves lower average application slowdown when three applications run concurrently.

Eight applications are included: Liblinear, XSBench, Graph500 BFS, GraphChi PageRank, Spark-KMeans, Metis PVC, Cassandra, and Memcached. SPEC CPU2017 gcc is omitted because it requires a separate licence. Figure 9 covers Mix1–4.

Layout

chameleon-ae/
├── README.md, LICENSE
├── run-all.sh                 # Figures 7/8, then Figure 9; parse and draw
├── fig78.sh, fig9.sh           # Independent evaluation entry points
├── scripts/                   # Environment, build, install and Guest setup
├── docs/environment.md        # Full deployment instructions
├── hyperalloc-6.18/
│   ├── linux/                 # Complete Guest + Host kernel source
│   ├── configs/               # Separate Guest/Host build configurations
│   ├── qemu/, llfree-c/       # Modified QEMU and shared allocator
│   ├── hermit/                # Guest module, memory server and protocol
│   ├── scripts/, tests/       # Build, deployment, VM tools and checks
│   └── build/                 # Generated locally; not distributed
├── benchmarks/
│   ├── apps/, vendor/         # Application and library sources/distributions
│   ├── dataset/               # Included input data
│   ├── scripts/, patches/     # Application launch and preparation scripts
│   └── SOURCE-MANIFEST.tsv    # Upstream revisions and input sources
└── ae/
    ├── README.md              # Experiment protocols, baselines and claim assessment
    ├── config/                # Evaluation configurations and site configuration
    ├── scripts/               # Three-run evaluation, parsing and plotting
    ├── results_baselines/     # Author-provided pre-measured baselines
    └── results/              # Reviewer measurements and generated figures

Guest and Host share one Linux source tree, with different configurations and build directories. The installed releases are 6.18.0-chameleon-guest and 6.18.0-chameleon-host; deployment outputs are under hyperalloc-6.18/build/deploy-{guest,host,qemu}. QEMU is based on version 8.2.1.

Evaluation environment

The tested compute platform uses a two-socket Intel server with 48 physical cores, 256 GiB RAM, KVM/EPT, PEBS support, and a Mellanox mlx5 InfiniBand adapter with SR-IOV. A separate memory server provides native RDMA connectivity. Figure 9 needs three compute-side VFs in separate IOMMU groups, one per Guest; the memory server can share one native RDMA interface across three independent pools. The scripts assign different physical cores to VM vCPUs, QEMU service threads and Host workload generators.

Use Ubuntu 22.04 userspace, Python 3, NumPy/Matplotlib, a working /dev/kvm, IOMMU/VFIO support, and SSH public-key access to the memory server. Setup needs sudo; experiments use sudo for the QEMU/numad guard where applicable. Native kernel RDMA drivers are used. Allow ample space for kernel builds, a 350 GiB sparse template disk, prepared datasets and VM overlays; these generated files are substantially larger than the roughly 37 GiB source/data package.

On an already prepared evaluation server, skip installation and begin with the commands below. Coordinate your evaluation slot with the authors in advance. We recommend that only one reviewer use the platform at a time to avoid interference between experiments. On your own servers, follow the environment guide. The main setup sequence is:

bash scripts/setup-host.sh --apply
python3 scripts/build-system.py --jobs 16 --apply
bash scripts/install-host.sh --apply
# Select the new Host kernel at reboot; enable IOMMU as described in the guide.
# Configure the experiment user's KVM/VFIO access and unlimited memlock,
# then log in again; see docs/environment.md.

# Configure the memory-server SSH target, its RDMA IP and the three VF addresses.
python3 scripts/configure-site.py --plan
# Edit ae/config/host.example.json for your hardware, then:
python3 scripts/configure-site.py --write
python3 scripts/prepare-rdma-server.py --install-deps --apply
python3 scripts/create-template.py --apply
python3 scripts/deploy-benchmarks.py --apply

# If the three VFs have not been prepared, provision them while the template is OFF.
# Replace PF_BDF, PF_INTERFACE and REVIEWER_LOGIN with your dedicated IB PF/user.
sudo python3 benchmarks/scripts/prepare-chameleon-fig9.py \
  --pf PF_BDF --interface PF_INTERFACE --user REVIEWER_LOGIN \
  --example ae/config/host.json --inventory ae/config/host.json
sudo python3 benchmarks/scripts/prepare-chameleon-fig9.py \
  --pf PF_BDF --interface PF_INTERFACE --user REVIEWER_LOGIN \
  --example ae/config/host.json --inventory ae/config/host.json --apply

ae/config/host.json contains site-specific addresses, VF identifiers, Guest names and ports. It contains no passwords. The guide also covers VF creation, remote memlock limits, template preparation, and uploading/running commands with guestctl.py. Keep the prepared template stopped and unchanged while its experiment overlays exist.

Run the evaluation

A complete evaluation on the prepared platform is expected to take 22 hours; please reserve 24 hours. This estimate excludes initial installation, compilation and dataset preparation.

From the artifact root, first inspect the plan. This does not start VMs:

./run-all.sh --plan

Run all supported experiments and draw Figures 7, 8 and 9:

sudo -v
./run-all.sh --results ae/results/evaluation-01

run-all.sh runs Figure 7/8 before Figure 9. Each application point and each Mix runs three times.

Run a figure or a smaller application selection:

./fig78.sh --results ae/results/single-01
./fig78.sh --apps xsbench --results ae/results/xsbench-01
./fig9.sh --mixes mix1 --results ae/results/mix1-01

See the evaluation details for the single-VM and multi-VM protocols, provided baselines, and how to assess the two claims.

Results and plotting

For --results ae/results/evaluation-01, the principal outputs are:

ae/results/evaluation-01/
├── fig78/
│   ├── manifest.json, summary.json, all-local.json
│   ├── runs/, raw/            # Per-run parameters, logs and sampled counters
│   ├── fig7.svg, fig7.pdf
│   └── fig8.svg, fig8.pdf
└── fig9/
    ├── config.json, summary.json
    ├── raw/, raw-apps/        # Mix and per-application measurements
    └── fig9.svg, fig9.pdf

Reparse and redraw without rerunning applications:

./run-all.sh --parse-only --results ae/results/evaluation-01
# Or redraw a single already-parsed figure:
python3 ae/scripts/plot_figures.py --figure fig78 \
  --input ae/results/evaluation-01/fig78/summary.json \
  --output-dir ae/results/evaluation-01/fig78