Skip to content

Integer multi GPU full benchmarks #20

Integer multi GPU full benchmarks

Integer multi GPU full benchmarks #20

# Run all integer benchmarks on an instance with CUDA and return parsed results to Slab CI bot.
name: Integer multi GPU full benchmarks
on:
workflow_dispatch:
schedule:
# Weekly benchmarks will be triggered each Saturday at 1a.m.
- cron: '0 1 * * 6'
env:
CARGO_TERM_COLOR: always
RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
RUST_BACKTRACE: "full"
RUST_MIN_STACK: "8388608"
SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
jobs:
setup-instance:
name: Setup instance (cuda-integer-full-multi-gpu-benchmarks)
runs-on: ubuntu-latest
if: github.event_name != 'schedule' ||
(github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
outputs:
runner-name: ${{ steps.start-instance.outputs.label }}
steps:
- name: Start instance
id: start-instance
uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
with:
mode: start
github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
slab-url: ${{ secrets.SLAB_BASE_URL }}
job-secret: ${{ secrets.JOB_SECRET }}
backend: hyperstack
profile: multi-h100-nvlink
cuda-integer-full-multi-gpu-benchmarks:
name: Execute multi GPU integer benchmarks for all operations flavor
needs: setup-instance
runs-on: ${{ needs.setup-instance.outputs.runner-name }}
timeout-minutes: 1440 # 24 hours
continue-on-error: true
strategy:
fail-fast: false
max-parallel: 1
matrix:
command: [integer, integer_multi_bit]
op_flavor: [default, unchecked]
# explicit include-based build matrix, of known valid options
include:
- os: ubuntu-22.04
cuda: "12.2"
gcc: 11
env:
CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
CMAKE_VERSION: 3.29.6
steps:
# Mandatory on hyperstack since a bootable volume is not re-usable yet.
- name: Install dependencies
run: |
sudo apt update
sudo apt install -y checkinstall zlib1g-dev libssl-dev
wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
cd cmake-${{ env.CMAKE_VERSION }}
./bootstrap
make -j"$(nproc)"
sudo make install
- name: Checkout tfhe-rs repo with tags
uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
with:
fetch-depth: 0
- name: Get benchmark details
run: |
{
echo "BENCH_DATE=$(date --iso-8601=seconds)";
echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
echo "COMMIT_HASH=$(git describe --tags --dirty)";
} >> "${GITHUB_ENV}"
- name: Set up home
# "Install rust" step require root user to have a HOME directory which is not set.
run: |
echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
- name: Install rust
uses: dtolnay/rust-toolchain@21dc36fb71dd22e3317045c0c31a3f4249868b17
with:
toolchain: nightly
- name: Export CUDA variables
if: ${{ !cancelled() }}
run: |
{
echo "CUDA_PATH=$CUDA_PATH";
echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH";
echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc";
} >> "${GITHUB_ENV}"
echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
# Specify the correct host compilers
- name: Export gcc and g++ variables
if: ${{ !cancelled() }}
run: |
{
echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
} >> "${GITHUB_ENV}"
- name: Checkout Slab repo
uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
with:
repository: zama-ai/slab
path: slab
token: ${{ secrets.FHE_ACTIONS_TOKEN }}
- name: Check device is detected
if: ${{ !cancelled() }}
run: nvidia-smi
- name: Run benchmarks with AVX512
run: |
make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_${{ matrix.command }}_gpu
- name: Parse results
run: |
python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
--database tfhe_rs \
--hardware "n3-H100x8-NVLink" \
--backend gpu \
--project-version "${{ env.COMMIT_HASH }}" \
--branch ${{ github.ref_name }} \
--commit-date "${{ env.COMMIT_DATE }}" \
--bench-date "${{ env.BENCH_DATE }}" \
--walk-subdirs \
--name-suffix avx512 \
--throughput
- name: Upload parsed results artifact
uses: actions/upload-artifact@0b2256b8c012f0828dc542b3febcab082c67f72b
with:
name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}
path: ${{ env.RESULTS_FILENAME }}
- name: Send data to Slab
shell: bash
run: |
echo "Computing HMac on results file"
SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
echo "Sending results to Slab..."
curl -v -k \
-H "Content-Type: application/json" \
-H "X-Slab-Repository: ${{ github.repository }}" \
-H "X-Slab-Command: store_data_v2" \
-H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-d @${{ env.RESULTS_FILENAME }} \
${{ secrets.SLAB_URL }}
slack-notify:
name: Slack Notification
needs: [ setup-instance, cuda-integer-full-multi-gpu-benchmarks ]
runs-on: ubuntu-latest
if: ${{ !success() && !cancelled() }}
continue-on-error: true
steps:
- name: Send message
uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
env:
SLACK_COLOR: ${{ needs.cuda-integer-full-multi-gpu-benchmarks.result }}
SLACK_MESSAGE: "Integer GPU full benchmarks finished with status: ${{ needs.cuda-integer-full-multi-gpu-benchmarks.result }}. (${{ env.ACTION_RUN_URL }})"
teardown-instance:
name: Teardown instance (cuda-integer-full-multi-gpu-benchmarks)
if: ${{ always() && needs.setup-instance.result != 'skipped' }}
needs: [ setup-instance, cuda-integer-full-multi-gpu-benchmarks ]
runs-on: ubuntu-latest
steps:
- name: Stop instance
id: stop-instance
uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
with:
mode: stop
github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
slab-url: ${{ secrets.SLAB_BASE_URL }}
job-secret: ${{ secrets.JOB_SECRET }}
label: ${{ needs.setup-instance.outputs.runner-name }}
- name: Slack Notification
if: ${{ failure() }}
continue-on-error: true
uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
env:
SLACK_COLOR: ${{ job.status }}
SLACK_MESSAGE: "Instance teardown (cuda-integer-full-multi-gpu-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"