microsoft · arrayka · Mar 30, 2026 · May 11, 2026 · May 11, 2026 · May 11, 2026
diff --git a/.github/actions/setup-disk-benchmark/action.yml b/.github/actions/setup-disk-benchmark/action.yml
@@ -17,14 +17,34 @@ inputs:
   extract-to:
     description: 'Directory to extract the dataset into'
     required: true
+  nvme-mount-path:
+    description: 'Mount point for the high-speed NVMe SSD'
+    required: false
+    default: '/mnt/nvme'
 
 runs:
   using: composite
   steps:
     - name: Install Rust
+      shell: bash
+      run: |
+        if ! command -v rustup &>/dev/null; then
+          curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
+          echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
+        fi
+
+    - name: Show Rust toolchain
       shell: bash
       run: rustup show
 
+    - name: Mount high-speed NVMe SSD
+      shell: bash
+      run: |
+        sudo mkdir -p -m 777 ${{ inputs.nvme-mount-path }}
+        sudo lsblk
+        sudo mkfs.ext4 /dev/nvme0n1
+        sudo mount /dev/nvme0n1 ${{ inputs.nvme-mount-path }}
+
     - name: Install system dependencies
       shell: bash
       run: |

diff --git a/.github/workflows/disk-benchmarks-aa.yml b/.github/workflows/disk-benchmarks-aa.yml
@@ -24,6 +24,8 @@ concurrency:
 env:
   RUST_BACKTRACE: 1
   PERF_INPUTS: diskann-benchmark/perf_test_inputs
+  NVME_MOUNT_PATH: /mnt/nvme
+  DISKANN_RUST: /mnt/nvme/diskann_rust
 
 defaults:
   run:
@@ -37,7 +39,7 @@ jobs:
   # A/A benchmark: run main vs main to detect environment noise.
   aa-benchmark:
     name: A/A - ${{ matrix.dataset }}
-    runs-on: ubuntu-latest
+    runs-on: [ self-hosted, 1ES.Pool=diskann-github, ubuntu-latest ]
     timeout-minutes: 120
     strategy:
       fail-fast: false
@@ -51,41 +53,49 @@ jobs:
             archive: openai-100K.tar.gz
 
     steps:
-      - name: Checkout main
+      # Shallow checkout to $GITHUB_WORKSPACE so the composite action is accessible via uses:.
+      - name: Checkout action definitions
         uses: actions/checkout@v4
         with:
-          ref: main
-          path: diskann_rust
-          lfs: true
+          sparse-checkout: .github
+          path: github_actions
 
       - name: Setup benchmark environment
-        uses: ./diskann_rust/.github/actions/setup-disk-benchmark
+        uses: ./github_actions/.github/actions/setup-disk-benchmark
         with:
           dataset: ${{ matrix.dataset }}
           archive: ${{ matrix.archive }}
-          extract-to: diskann_rust/target/tmp
+          extract-to: ${{ env.DISKANN_RUST }}/target/tmp
+          nvme-mount-path: ${{ env.NVME_MOUNT_PATH }}
+
+      - name: Checkout main
+        uses: actions/checkout@v4
+        with:
+          ref: main
+          path: ${{ env.DISKANN_RUST }}
+          lfs: true
 
       # A/A: build once, run twice (identical code — only detecting environment noise)
       - name: Build benchmark binary
-        working-directory: diskann_rust
+        working-directory: ${{ env.DISKANN_RUST }}
         run: cargo build -p diskann-benchmark --features disk-index --release
 
       - name: Run baseline benchmark
-        working-directory: diskann_rust
+        working-directory: ${{ env.DISKANN_RUST }}
         run: |
           cargo run -p diskann-benchmark --features disk-index --release -- \
             run --input-file ${{ env.PERF_INPUTS }}/${{ matrix.config }} \
             --output-file target/tmp/${{ matrix.dataset }}_baseline.json
 
       - name: Run target benchmark
-        working-directory: diskann_rust
+        working-directory: ${{ env.DISKANN_RUST }}
         run: |
           cargo run -p diskann-benchmark --features disk-index --release -- \
             run --input-file ${{ env.PERF_INPUTS }}/${{ matrix.config }} \
             --output-file target/tmp/${{ matrix.dataset }}_target.json
 
       - name: Validate benchmark results
-        working-directory: diskann_rust
+        working-directory: ${{ env.DISKANN_RUST }}
         run: |
           cargo run -p diskann-benchmark --features disk-index --release -- \
             check run \
@@ -100,8 +110,8 @@ jobs:
         with:
           name: aa-results-${{ matrix.dataset }}
           path: |
-            diskann_rust/target/tmp/${{ matrix.dataset }}_target.json
-            diskann_rust/target/tmp/${{ matrix.dataset }}_baseline.json
+            ${{ env.DISKANN_RUST }}/target/tmp/${{ matrix.dataset }}_target.json
+            ${{ env.DISKANN_RUST }}/target/tmp/${{ matrix.dataset }}_baseline.json
           retention-days: 30
 
   # Notify diskann-disk-maintainers on A/A failure — but only when the failure

diff --git a/.github/workflows/disk-benchmarks.yml b/.github/workflows/disk-benchmarks.yml
@@ -42,6 +42,9 @@ concurrency:
 env:
   RUST_BACKTRACE: 1
   PERF_INPUTS: diskann-benchmark/perf_test_inputs
+  NVME_MOUNT_PATH: /mnt/nvme
+  DISKANN_RUST: /mnt/nvme/diskann_rust
+  BASELINE: /mnt/nvme/baseline
 
 defaults:
   run:
@@ -54,7 +57,7 @@ jobs:
   # Macro benchmark: compare current branch against baseline
   macro-benchmark:
     name: Macro Benchmark - ${{ matrix.dataset }}
-    runs-on: ubuntu-latest
+    runs-on: [ self-hosted, 1ES.Pool=diskann-github, ubuntu-latest ]
     # TODO: For production benchmarks, consider using a self-hosted runner with:
     # - NVMe storage for consistent I/O performance
     # - CPU pinning (taskset) for reduced variance
@@ -74,53 +77,61 @@ jobs:
             data_dir: OpenAIArXiv
 
     steps:
+      # Shallow checkout to $GITHUB_WORKSPACE so the composite action is accessible via uses:.
+      - name: Checkout action definitions
+        uses: actions/checkout@v4
+        with:
+          sparse-checkout: .github
+          path: github_actions
+
+      - name: Setup benchmark environment
+        uses: ./github_actions/.github/actions/setup-disk-benchmark
+        with:
+          dataset: ${{ matrix.dataset }}
+          archive: ${{ matrix.archive }}
+          extract-to: ${{ env.DISKANN_RUST }}/target/tmp
+          nvme-mount-path: ${{ env.NVME_MOUNT_PATH }}
+
       - name: Checkout current branch
         uses: actions/checkout@v4
         with:
-          path: diskann_rust
+          path: ${{ env.DISKANN_RUST }}
           lfs: true
 
       - name: Checkout baseline (${{ inputs.baseline_ref || 'main' }})
         uses: actions/checkout@v4
         with:
           ref: ${{ inputs.baseline_ref || 'main' }}
-          path: baseline
+          path: ${{ env.BASELINE }}
           lfs: true
 
-      - name: Setup benchmark environment
-        uses: ./diskann_rust/.github/actions/setup-disk-benchmark
-        with:
-          dataset: ${{ matrix.dataset }}
-          archive: ${{ matrix.archive }}
-          extract-to: diskann_rust/target/tmp
-
       - name: Copy dataset to baseline
         run: |
-          mkdir -p baseline/target/tmp
-          cp -r diskann_rust/target/tmp/${{ matrix.data_dir }} baseline/target/tmp/
+          mkdir -p ${{ env.BASELINE }}/target/tmp
+          cp -r ${{ env.DISKANN_RUST }}/target/tmp/${{ matrix.data_dir }} ${{ env.BASELINE }}/target/tmp/
 
       - name: Run baseline benchmark
-        working-directory: baseline
+        working-directory: ${{ env.BASELINE }}
         run: |
           cargo run -p diskann-benchmark --features disk-index --release -- \
-            run --input-file ../diskann_rust/${{ env.PERF_INPUTS }}/${{ matrix.config }} \
+            run --input-file ${{ env.DISKANN_RUST }}/${{ env.PERF_INPUTS }}/${{ matrix.config }} \
             --output-file target/tmp/${{ matrix.dataset }}_baseline.json
 
       - name: Run current branch benchmark
-        working-directory: diskann_rust
+        working-directory: ${{ env.DISKANN_RUST }}
         run: |
           cargo run -p diskann-benchmark --features disk-index --release -- \
             run --input-file ${{ env.PERF_INPUTS }}/${{ matrix.config }} \
             --output-file target/tmp/${{ matrix.dataset }}_target.json
 
       - name: Validate benchmark results
-        working-directory: diskann_rust
+        working-directory: ${{ env.DISKANN_RUST }}
         run: |
           cargo run -p diskann-benchmark --features disk-index --release -- \
             check run \
             --tolerances ${{ env.PERF_INPUTS }}/disk-index-tolerances.json \
             --input-file ${{ env.PERF_INPUTS }}/${{ matrix.config }} \
-            --before ../baseline/target/tmp/${{ matrix.dataset }}_baseline.json \
+            --before ${{ env.BASELINE }}/target/tmp/${{ matrix.dataset }}_baseline.json \
             --after target/tmp/${{ matrix.dataset }}_target.json
 
       - name: Upload benchmark results
@@ -129,6 +140,6 @@ jobs:
         with:
           name: benchmark-results-${{ matrix.dataset }}
           path: |
-            diskann_rust/target/tmp/${{ matrix.dataset }}_target.json
-            baseline/target/tmp/${{ matrix.dataset }}_baseline.json
+            ${{ env.DISKANN_RUST }}/target/tmp/${{ matrix.dataset }}_target.json
+            ${{ env.BASELINE }}/target/tmp/${{ matrix.dataset }}_baseline.json
           retention-days: 30