[EM] Small tests for the no-concat option. (#11566)

- Update document for the new adaptive cache.

Author: trivialfis

python-package/xgboost/core.py

@@ -540,7 +540,7 @@ class DataIter(ABC):  # pylint: disable=too-many-instance-attributes
         with GPU-based :py:class:`ExtMemQuantileDMatrix`. When using GPU-based external
         memory with the data cached in the host memory, XGBoost can concatenate the
         pages internally to increase the batch size for the GPU. The default page size
-        is about 1/8 of the total device memory. Users can manually set the value based
+        is about 1/16 of the total device memory. Users can manually set the value based
         on the actual hardware and datasets. Set this to 0 to disable page
         concatenation.
 

src/data/data.cc

@@ -35,7 +35,6 @@
 #include "../data/iterative_dmatrix.h"        // for IterativeDMatrix
 #include "./sparse_page_dmatrix.h"            // for SparsePageDMatrix
 #include "array_interface.h"                  // for ArrayInterfaceHandler, ArrayInterface, Dispa...
-#include "batch_utils.h"                      // for MatchingPageBytes
 #include "cat_container.h"                    // for CatContainer
 #include "dmlc/base.h"                        // for BeginPtr
 #include "dmlc/common.h"                      // for OMPException

tests/cpp/data/test_ellpack_page_raw_format.cu

@@ -198,7 +198,11 @@ TEST(EllpackPageRawFormat, DevicePageConcat) {
       EXPECT_TRUE(page.Impl()->IsDense());
       CalcCacheMapping(&ctx, page.Impl()->IsDense(), cuts, min_cache_page_bytes, ext_info, false,
                        &cinfo);
-      EXPECT_EQ(cinfo.buffer_rows.size(), 4ul);
+      if (min_cache_page_bytes == ::xgboost::cuda_impl::MatchingPageBytes()) {
+        EXPECT_EQ(cinfo.NumBatchesCc(), ext_info.n_batches);
+      } else {
+        EXPECT_EQ(cinfo.buffer_rows.size(), 4ul);
+      }
       policy.SetCuts(page.Impl()->CutsShared(), ctx.Device(), std::move(cinfo));
     }
 
@@ -219,6 +223,11 @@ TEST(EllpackPageRawFormat, DevicePageConcat) {
     return mem_cache;
   };
 
+  {
+    auto mem_cache =
+        test(::xgboost::cuda_impl::MatchingPageBytes(), ::xgboost::cuda_impl::AutoHostRatio());
+    ASSERT_EQ(mem_cache->d_pages.size(), 8);
+  }
   {
     auto mem_cache = test(n_features * n_samples, ::xgboost::cuda_impl::AutoHostRatio());
     ASSERT_EQ(mem_cache->h_pages.size(), 4);

tests/cpp/data/test_extmem_quantile_dmatrix.cu

@@ -74,7 +74,7 @@ class EllpackHostCacheTest : public ::testing::TestWithParam<std::tuple<double,
     auto p_fmat = RandomDataGenerator{NumSamples(), NumFeatures(), sparsity}
                       .Device(ctx.Device())
                       .GenerateDMatrix();
-    bst_idx_t min_page_cache_bytes = 0;
+    bst_idx_t min_page_cache_bytes = ::xgboost::cuda_impl::MatchingPageBytes();
     if (is_concat) {
       min_page_cache_bytes =
           p_fmat->GetBatches<EllpackPage>(&ctx, param).begin().Page()->Impl()->MemCostBytes() / 3;
@@ -111,9 +111,8 @@ class EllpackHostCacheTest : public ::testing::TestWithParam<std::tuple<double,
 };
 
 TEST_P(EllpackHostCacheTest, Basic) {
-  auto ctx = MakeCUDACtx(0);
-  auto [sparsity, min_page_cache_bytes, cache_host_ratio] = this->GetParam();
-  this->Run(sparsity, min_page_cache_bytes, cache_host_ratio);
+  auto [sparsity, is_concat, cache_host_ratio] = this->GetParam();
+  this->Run(sparsity, is_concat, cache_host_ratio);
 }
 
 INSTANTIATE_TEST_SUITE_P(