Add environment variables for I/O thread pool and slice sizes

cpp/src/io/utilities/config_utils.cpp

@@ -42,7 +42,7 @@ enum class usage_policy : uint8_t { OFF, GDS, ALWAYS };
  */
 usage_policy get_env_policy()
 {
-  static auto const env_val = getenv_or("LIBCUDF_CUFILE_POLICY", "GDS");
+  static auto const env_val = getenv_or<std::string>("LIBCUDF_CUFILE_POLICY", "GDS");
   if (env_val == "OFF") return usage_policy::OFF;
   if (env_val == "GDS") return usage_policy::GDS;
   if (env_val == "ALWAYS") return usage_policy::ALWAYS;
@@ -69,7 +69,7 @@ enum class usage_policy : uint8_t { OFF, STABLE, ALWAYS };
  */
 usage_policy get_env_policy()
 {
-  static auto const env_val = getenv_or("LIBCUDF_NVCOMP_POLICY", "STABLE");
+  static auto const env_val = getenv_or<std::string>("LIBCUDF_NVCOMP_POLICY", "STABLE");
   if (env_val == "OFF") return usage_policy::OFF;
   if (env_val == "STABLE") return usage_policy::STABLE;
   if (env_val == "ALWAYS") return usage_policy::ALWAYS;

cpp/src/io/utilities/config_utils.hpp

@@ -15,6 +15,7 @@
  */
 #pragma once
 
+#include <sstream>
 #include <string>
 
 namespace cudf::io::detail {
@@ -23,7 +24,17 @@ namespace cudf::io::detail {
  * @brief Returns the value of the environment variable, or a default value if the variable is not
  * present.
  */
-std::string getenv_or(std::string const& env_var_name, std::string_view default_val);
+template <typename T>
+T getenv_or(std::string_view env_var_name, T default_val)
+{
+  auto const env_val = std::getenv(env_var_name.data());
+  if (env_val == nullptr) { return default_val; }
+
+  std::stringstream sstream(env_val);
+  T converted_val;
+  sstream >> converted_val;
+  return converted_val;
+}
 
 namespace cufile_integration {
 

cpp/src/io/utilities/file_io_utilities.cpp

@@ -91,7 +91,7 @@ void cufile_shim::modify_cufile_json() const
   temp_directory tmp_config_dir{"cudf_cufile_config"};
 
   // Modify the config file based on the policy
-  auto const config_file_path = getenv_or(json_path_env_var, "/etc/cufile.json");
+  auto const config_file_path = getenv_or<std::string>(json_path_env_var, "/etc/cufile.json");
   std::ifstream user_config_file(config_file_path);
   // Modified config file is stored in a temporary directory
   auto const cudf_config_path = tmp_config_dir.path() + "/cufile.json";
@@ -170,7 +170,8 @@ cufile_registered_file::~cufile_registered_file() { shim->handle_deregister(cf_h
 cufile_input_impl::cufile_input_impl(std::string const& filepath)
   : shim{cufile_shim::instance()},
     cf_file(shim, filepath, O_RDONLY | O_DIRECT),
-    pool(16)  // The benefit from multithreaded read plateaus around 16 threads
+    // The benefit from multithreaded read plateaus around 16 threads
+    pool(getenv_or("LIBCUDF_CUFILE_THREAD_COUNT", 16))
 {
   pool.sleep_duration = 10;
 }
@@ -194,9 +195,11 @@ std::vector<std::future<ResultT>> make_sliced_tasks(
   F function, DataT* ptr, size_t offset, size_t size, cudf::detail::thread_pool& pool)
 {
   std::vector<std::future<ResultT>> slice_tasks;
-  constexpr size_t max_slice_bytes = 4 * 1024 * 1024;
-  size_t const n_slices            = util::div_rounding_up_safe(size, max_slice_bytes);
-  size_t slice_offset              = 0;
+  constexpr size_t default_max_slice_bytes = 4 * 1024 * 1024;
+  static auto const max_slice_bytes =
+    getenv_or("LIBCUDF_CUFILE_SLICE_SIZE", default_max_slice_bytes);
+  size_t const n_slices = util::div_rounding_up_safe(size, max_slice_bytes);
+  size_t slice_offset   = 0;
   for (size_t t = 0; t < n_slices; ++t) {
     DataT* ptr_slice = ptr + slice_offset;
 
@@ -250,7 +253,7 @@ size_t cufile_input_impl::read(size_t offset,
 cufile_output_impl::cufile_output_impl(std::string const& filepath)
   : shim{cufile_shim::instance()},
     cf_file(shim, filepath, O_CREAT | O_RDWR | O_DIRECT, 0664),
-    pool(16)
+    pool(getenv_or("LIBCUDF_CUFILE_THREAD_COUNT", 16))
 {
 }
 

docs/cudf/source/basics/io-gds-integration.rst

@@ -31,3 +31,9 @@ Operations that support the use of GPUDirect Storage:
 - `to_csv`
 - `to_parquet`
 - `to_orc`
+
+Several parameters that can be used to tune the performance of GDS-enabled I/O are exposed through environment variables:
+
+- ``LIBCUDF_CUFILE_THREAD_COUNT``: Integral value, maximum number of parallel reads/writes per file (default 16);
+- ``LIBCUDF_CUFILE_SLICE_SIZE``: Integral value, maximum size of each GDS read/write, in bytes (default 4MB).
+  Larger I/O operations are split into multiple calls.