[graphbolt] Disable liburing (#7366)

CMakeLists.txt

@@ -20,6 +20,7 @@ dgl_option(BUILD_TYPE "Type of the build: dev, dogfood or release" "dev")
 message(STATUS "Build for ${BUILD_TYPE}")
 
 dgl_option(USE_CUDA "Build with CUDA" OFF)
+dgl_option(USE_LIBURING "Build with liburing" OFF)
 dgl_option(TORCH_PYTHON_INTERPS "Python interpreter for building sub-components" python3)
 
 # Conda build related options.
@@ -349,23 +350,26 @@ endif(EXTERNAL_PHMAP_PATH)
 target_include_directories(dgl PRIVATE "tensoradapter/include")
 target_include_directories(dgl PRIVATE "third_party/pcg/include")
 
-if(CMAKE_SYSTEM_NAME MATCHES "Linux")
-  include(ExternalProject)
-  set(LIBURING_INSTALL_DIR ${CMAKE_BINARY_DIR}/third_party/liburing)
-  ExternalProject_Add(
-    liburing
-    SOURCE_DIR ${CMAKE_SOURCE_DIR}/third_party/liburing
-    CONFIGURE_COMMAND <SOURCE_DIR>/configure --prefix=/
-    BUILD_COMMAND bash -c "make -j 4"
-    BUILD_IN_SOURCE ON
-    INSTALL_COMMAND make install DESTDIR=${LIBURING_INSTALL_DIR}
-    BUILD_BYPRODUCTS ${LIBURING_INSTALL_DIR}/lib/liburing.a
-    BUILD_BYPRODUCTS ${LIBURING_INSTALL_DIR}/include
-    DOWNLOAD_EXTRACT_TIMESTAMP true
-  )
-  set(LIBURING_INCLUDE ${LIBURING_INSTALL_DIR}/include)
-  set(LIBURING ${LIBURING_INSTALL_DIR}/lib/liburing.a)
-endif()
+if(USE_LIBURING)
+  if(CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_definitions(-DHAVE_LIBRARY_LIBURING)
+    include(ExternalProject)
+    set(LIBURING_INSTALL_DIR ${CMAKE_BINARY_DIR}/third_party/liburing)
+    ExternalProject_Add(
+      liburing
+      SOURCE_DIR ${CMAKE_SOURCE_DIR}/third_party/liburing
+      CONFIGURE_COMMAND <SOURCE_DIR>/configure --prefix=/
+      BUILD_COMMAND bash -c "make -j 4"
+      BUILD_IN_SOURCE ON
+      INSTALL_COMMAND make install DESTDIR=${LIBURING_INSTALL_DIR}
+      BUILD_BYPRODUCTS ${LIBURING_INSTALL_DIR}/lib/liburing.a
+      BUILD_BYPRODUCTS ${LIBURING_INSTALL_DIR}/include
+      DOWNLOAD_EXTRACT_TIMESTAMP true
+    )
+    set(LIBURING_INCLUDE ${LIBURING_INSTALL_DIR}/include)
+    set(LIBURING ${LIBURING_INSTALL_DIR}/lib/liburing.a)
+  endif()
+endif(USE_LIBURING)
 
 if(EXTERNAL_NANOFLANN_PATH)
   include_directories(SYSTEM ${EXTERNAL_NANOFLANN_PATH})
@@ -588,7 +592,9 @@ if(BUILD_GRAPHBOLT)
   if(USE_CUDA)
     add_dependencies(graphbolt gpu_cache)
   endif(USE_CUDA)
-  if(CMAKE_SYSTEM_NAME MATCHES "Linux")
-    add_dependencies(graphbolt liburing)
-  endif()
+  if(USE_LIBURING)
+    if(CMAKE_SYSTEM_NAME MATCHES "Linux")
+      add_dependencies(graphbolt liburing)
+    endif()
+  endif(USE_LIBURING)
 endif(BUILD_GRAPHBOLT)

graphbolt/CMakeLists.txt

@@ -76,9 +76,11 @@ include_directories(BEFORE ${BOLT_DIR}
                            "../third_party/pcg/include")
 target_link_libraries(${LIB_GRAPHBOLT_NAME} "${TORCH_LIBRARIES}")
 if(CMAKE_SYSTEM_NAME MATCHES "Linux")
-  target_include_directories(${LIB_GRAPHBOLT_NAME} PRIVATE "../third_party/liburing/src/include")
-  get_filename_component(PARENT_DIR "${CMAKE_SOURCE_DIR}" DIRECTORY)
-  target_link_libraries(${LIB_GRAPHBOLT_NAME} ${PARENT_DIR}/build/third_party/liburing/lib/liburing.a)
+  if(USE_LIBURING)
+    target_include_directories(${LIB_GRAPHBOLT_NAME} PRIVATE "../third_party/liburing/src/include")
+    get_filename_component(PARENT_DIR "${CMAKE_SOURCE_DIR}" DIRECTORY)
+    target_link_libraries(${LIB_GRAPHBOLT_NAME} ${PARENT_DIR}/build/third_party/liburing/lib/liburing.a)
+  endif(USE_LIBURING)
 endif()
 
 if(USE_CUDA)

graphbolt/src/cnumpy.cc

@@ -20,7 +20,11 @@ static constexpr int kDiskAlignmentSize = 4096;
 OnDiskNpyArray::OnDiskNpyArray(
     std::string filename, torch::ScalarType dtype, torch::Tensor shape)
     : filename_(filename), dtype_(dtype) {
-#ifdef __linux__
+#ifndef __linux__
+  throw std::runtime_error(
+      "OnDiskNpyArray is not supported on non-Linux systems.");
+#endif
+#ifdef HAVE_LIBRARY_LIBURING
   ParseNumpyHeader(shape);
   file_description_ = open(filename.c_str(), O_RDONLY | O_DIRECT);
   if (file_description_ == -1) {
@@ -35,7 +39,9 @@ OnDiskNpyArray::OnDiskNpyArray(
   for (int64_t t = 0; t < num_thread_; t++) {
     io_uring_queue_init(group_size_, &io_uring_queue_[t], 0);
   }
-#endif  // __linux__
+#else
+  throw std::runtime_error("DiskBasedFeature is not available now.");
+#endif  // HAVE_LIBRARY_LIBURING
 }
 
 c10::intrusive_ptr<OnDiskNpyArray> OnDiskNpyArray::Create(
@@ -44,17 +50,17 @@ c10::intrusive_ptr<OnDiskNpyArray> OnDiskNpyArray::Create(
 }
 
 OnDiskNpyArray::~OnDiskNpyArray() {
-#ifdef __linux__
+#ifdef HAVE_LIBRARY_LIBURING
   // IO queue exit.
   for (int64_t t = 0; t < num_thread_; t++) {
     io_uring_queue_exit(&io_uring_queue_[t]);
   }
   close(file_description_);
-#endif  // __linux__
+#endif  // HAVE_LIBRARY_LIBURING
 }
 
 void OnDiskNpyArray::ParseNumpyHeader(torch::Tensor shape) {
-#ifdef __linux__
+#ifdef HAVE_LIBRARY_LIBURING
   // Parse numpy file header to get basic info of feature.
   size_t word_size = c10::elementSize(dtype_);
   int64_t num_dim = shape.numel();
@@ -80,19 +86,19 @@ void OnDiskNpyArray::ParseNumpyHeader(torch::Tensor shape) {
   // Get prefix length for computing feature offset,
   // add one for new-line character.
   prefix_len_ = header.size() + 1;
-#endif  // __linux__
+#endif  // HAVE_LIBRARY_LIBURING
 }
 
 torch::Tensor OnDiskNpyArray::IndexSelect(torch::Tensor index) {
-#ifdef __linux__
+#ifdef HAVE_LIBRARY_LIBURING
   return IndexSelectIOUring(index);
 #else
   TORCH_CHECK(false, "OnDiskNpyArray is not supported on non-Linux systems.");
   return torch::empty({0});
-#endif  // __linux__
+#endif  // HAVE_LIBRARY_LIBURING
 }
 
-#ifdef __linux__
+#ifdef HAVE_LIBRARY_LIBURING
 torch::Tensor OnDiskNpyArray::IndexSelectIOUring(torch::Tensor index) {
   index = index.to(torch::kLong);
   // The minimum page size to contain one feature.
@@ -215,7 +221,6 @@ torch::Tensor OnDiskNpyArray::IndexSelectIOUring(torch::Tensor index) {
 
   return result;
 }
-#endif  // __linux__
-
+#endif  // HAVE_LIBRARY_LIBURING
 }  // namespace storage
 }  // namespace graphbolt

graphbolt/src/cnumpy.h

@@ -9,10 +9,10 @@
 #include <stdlib.h>
 #include <torch/script.h>
 
-#ifdef __linux__
+#ifdef HAVE_LIBRARY_LIBURING
 #include <liburing.h>
 #include <unistd.h>
-#endif
+#endif  // HAVE_LIBRARY_LIBURING
 
 #include <cassert>
 #include <cstdio>
@@ -60,7 +60,7 @@ class OnDiskNpyArray : public torch::CustomClassHolder {
    */
   torch::Tensor IndexSelect(torch::Tensor index);
 
-#ifdef __linux__
+#ifdef HAVE_LIBRARY_LIBURING
   /**
    * @brief Index-select operation on an on-disk numpy array using IO Uring for
    * asynchronous I/O.
@@ -76,8 +76,7 @@ class OnDiskNpyArray : public torch::CustomClassHolder {
    * @throws std::runtime_error If index is out of range.
    */
   torch::Tensor IndexSelectIOUring(torch::Tensor index);
-#endif  // __linux__
-
+#endif  // HAVE_LIBRARY_LIBURING
  private:
   std::string filename_;              // Path to numpy file.
   int file_description_;              // File description.
@@ -88,9 +87,9 @@ class OnDiskNpyArray : public torch::CustomClassHolder {
   int num_thread_;                    // Default thread number.
   int64_t group_size_ = 512;          // Default group size.
 
-#ifdef __linux__
+#ifdef HAVE_LIBRARY_LIBURING
   io_uring* io_uring_queue_;  // io_uring queue.
-#endif
+#endif                        // HAVE_LIBRARY_LIBURING
 };
 
 }  // namespace storage

tests/python/pytorch/graphbolt/impl/test_disk_based_feature_store.py

@@ -18,9 +18,8 @@ def to_on_disk_numpy(test_dir, name, t):
 
 
 @unittest.skipIf(
-    sys.platform.startswith("win"),
-    reason="Tests for disk dataset can only deployed on Linux,"
-    "because the io_uring is only supportted by Linux kernel.",
+    True,
+    reason="DiskBasedFeature is not available now.",
 )
 def test_disk_based_feature():
     with tempfile.TemporaryDirectory() as test_dir:
@@ -85,6 +84,10 @@ def test_disk_based_feature():
     reason="Tests for disk dataset can only deployed on Linux,"
     "because the io_uring is only supportted by Linux kernel.",
 )
+@unittest.skipIf(
+    True,
+    reason="DiskBasedFeature is not available now.",
+)
 @pytest.mark.parametrize(
     "dtype",
     [
@@ -129,6 +132,10 @@ def test_more_disk_based_feature(dtype, idtype, shape, index):
     reason="Tests for large disk dataset can only deployed on Linux,"
     "because the io_uring is only supportted by Linux kernel.",
 )
+@unittest.skipIf(
+    True,
+    reason="DiskBasedFeature is not available now.",
+)
 def test_disk_based_feature_repr():
     with tempfile.TemporaryDirectory() as test_dir:
         a = torch.tensor([[1, 2, 3], [4, 5, 6]])