Explicit build dependency on nvidia_peermem (#201)

.github/workflows/codeql.yml → .github/workflows/codeql-analysis.yml

@@ -2,8 +2,9 @@ name: "CodeQL"
 
 on:
   push:
-    branches:
-      - '**'
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
   schedule:
     - cron: "30 1 * * 1"
 
@@ -42,8 +43,10 @@ jobs:
       run: |
         git config --global --add safe.directory /__w/mscclpp/mscclpp
 
-    - name: Autobuild
-      uses: github/codeql-action/autobuild@v2
+    - name: Build
+      run: |
+        MPI_HOME=/usr/local/mpi cmake -DBYPASS_PEERMEM_CHECK=ON .
+        make -j
 
     - name: Perform CodeQL Analysis
       uses: github/codeql-action/analyze@v2

CMakeLists.txt

@@ -26,6 +26,7 @@ option(USE_NPKIT "Use NPKIT" ON)
 option(BUILD_TESTS "Build tests" ON)
 option(BUILD_PYTHON_BINDINGS "Build Python bindings" ON)
 option(ALLOW_GDRCOPY "Use GDRCopy, if available" OFF)
+option(BYPASS_PEERMEM_CHECK "Bypass checking nvidia_peermem" OFF)
 
 # Find CUDAToolkit. Set CUDA flags based on the detected CUDA version
 find_package(CUDAToolkit REQUIRED)
@@ -44,6 +45,16 @@ if(CUDAToolkit_FOUND)
 endif()
 set(CUDA_LIBRARIES CUDA::cudart CUDA::cuda_driver)
 
+# Find if nvidia_peermem is installed and loaded
+if(NOT BYPASS_PEERMEM_CHECK)
+    execute_process(COMMAND sh -c "lsmod | grep nvidia_peermem"
+                    RESULT_VARIABLE lsmod_result
+                    OUTPUT_VARIABLE lsmod_output)
+    if(NOT lsmod_result EQUAL 0)
+        message(FATAL_ERROR "nvidia_peermem is not installed or not loaded.")
+    endif()
+endif()
+
 # Find ibverbs and libnuma
 find_package(IBVerbs REQUIRED)
 find_package(NUMA REQUIRED)

docs/quickstart.md

@@ -11,7 +11,12 @@
     * NVIDIA A100 GPUs + CUDA >= 11.1.1
     * NVIDIA H100 GPUs + CUDA >= 12.0.0
 * OS: tested over Ubuntu 18.04 and 20.04
-* Libraries: [libnuma](https://github.com/numactl/numactl), [GDRCopy](https://github.com/NVIDIA/gdrcopy) (optional), MPI (optional)
+* Libraries: [libnuma](https://github.com/numactl/numactl), MPI (optional)
+* Others
+    * `nvidia_peermem` driver should be loaded on all nodes. Check it via:
+        ```
+        lsmod | grep nvidia_peermem
+        ```
 
 ## Build from Source