RowToColumnTests.Biggest OOM on A3000 Laptop GPU with 6 GiB

[gerashegalov]

Repro:

./build/build-in-docker clean install \
  -DGPU_ARCHS=NATIVE \
  -DBUILD_TESTS=ON  \
  -DskipTests  \
  -Dlibcudf.clean.skip=false

Log file contet

RowToColumnTests.Biggest
unknown file: Failure
C++ exception with description "std::bad_alloc: out_of_memory: CUDA error at: /home/user/gits/NVIDIA/spark-rapids-jni/target/libcudf-install/include/rmm/mr/device/cuda_memory_resource.hpp:60: cudaErrorMemoryAllocation out of memory" thrown in the test body.

The same test passes on Quadro RTX 6000 with 24GiB

The test hard-codes a pretty large table in absolute terms:

spark-rapids-jni/src/main/cpp/tests/row_conversion.cpp

Lines 412 to 418 in 6d68216

	constexpr auto num_rows = 2 * 1024 * 1024;
	for (int i = 0; i < 128; ++i) {
	cols.push_back(cudf::test::fixed_width_column_wrapper<int32_t>(r + num_rows * i,
	r + num_rows * i + num_rows));
	views.push_back(cols.back());
	schema.push_back(cudf::data_type{cudf::type_id::INT32});
	}

(2 * 1024 * 1024) rows * 128 cols * 4 = 1GiB

which is then replicated by conversion. Presumably the smaller GPU cannot pass this test as is.

The test should fail or be skipped with a clear message indicating that the GPU under test is too small.

[mattahrens]

Scope: update documentation related to test environment requirements for GPU memory.