wip.1206.1

apache · Dec 8, 2023 · d32f7f5 · d32f7f5
1 parent ddaa62a
commit d32f7f5
Show file tree

Hide file tree

Showing 15 changed files with 287 additions and 800 deletions.
diff --git a/...-clickhouse/src/main/scala/io/glutenproject/backendsapi/clickhouse/CHTransformerApi.scala b/...-clickhouse/src/main/scala/io/glutenproject/backendsapi/clickhouse/CHTransformerApi.scala
@@ -96,6 +96,12 @@ class CHTransformerApi extends TransformerApi with Logging {
           nativeConfMap.put(groupBySpillKey, groupBySpillValue.toLong.toString)
         }
 
+        val maxMemoryUsageKey = settingPrefix + "max_memory_usage";
+        if (!nativeConfMap.containsKey(maxMemoryUsageKey)) {
+          val maxMemoryUsageValue = offHeapSize
+          nativeConfMap.put(maxMemoryUsageKey, maxMemoryUsageValue.toLong.toString)
+        }
+
         // Only set default max_bytes_before_external_join for CH when join_algorithm is grace_hash
         val joinAlgorithmKey = settingPrefix + "join_algorithm";
         if (

diff --git a/...ends-clickhouse/src/main/scala/io/glutenproject/metrics/HashAggregateMetricsUpdater.scala b/...ends-clickhouse/src/main/scala/io/glutenproject/metrics/HashAggregateMetricsUpdater.scala
@@ -96,10 +96,12 @@ class HashAggregateMetricsUpdater(val metrics: Map[String, SQLMetric])
 object HashAggregateMetricsUpdater {
   val INCLUDING_PROCESSORS = Array(
     "AggregatingTransform",
+    "StreamingAggregatingTransform",
     "MergingAggregatedTransform",
-    "ScalableMergingAggregatedTransform")
+    "GraceMergingAggregatedTransform")
   val CH_PLAN_NODE_NAME = Array(
     "AggregatingTransform",
+    "StreamingAggregatingTransform",
     "MergingAggregatedTransform",
-    "ScalableMergingAggregatedTransform")
+    "GraceMergingAggregatedTransform")
 }
diff --git a/...scala/io/glutenproject/execution/GlutenClickHouseTPCHColumnarShuffleParquetAQESuite.scala b/...scala/io/glutenproject/execution/GlutenClickHouseTPCHColumnarShuffleParquetAQESuite.scala
@@ -66,8 +66,6 @@ class GlutenClickHouseTPCHColumnarShuffleParquetAQESuite
         assert(plans(2).metrics("outputRows").value === 600572)
 
         assert(plans(1).metrics("inputRows").value === 591673)
-        assert(plans(1).metrics("resizeInputRows").value === 4)
-        assert(plans(1).metrics("resizeOutputRows").value === 4)
         assert(plans(1).metrics("outputRows").value === 4)
         assert(plans(1).metrics("outputVectors").value === 1)
 
@@ -93,8 +91,6 @@ class GlutenClickHouseTPCHColumnarShuffleParquetAQESuite
           assert(plans(2).metrics("filesSize").value === 17777735)
 
           assert(plans(1).metrics("inputRows").value === 591673)
-          assert(plans(1).metrics("resizeInputRows").value === 4)
-          assert(plans(1).metrics("resizeOutputRows").value === 4)
           assert(plans(1).metrics("outputRows").value === 4)
           assert(plans(1).metrics("outputVectors").value === 1)
 

diff --git a/.../src/test/scala/io/glutenproject/execution/metrics/GlutenClickHouseTPCHMetricsSuite.scala b/.../src/test/scala/io/glutenproject/execution/metrics/GlutenClickHouseTPCHMetricsSuite.scala
@@ -311,15 +311,14 @@ class GlutenClickHouseTPCHMetricsSuite extends GlutenClickHouseTPCHAbstractSuite
         .get(0)
         .getProcessors
         .get(0)
-        .getInputRows == 591677)
-
+        .getInputRows == 591673)
     assert(
       nativeMetricsData.metricsDataList
         .get(4)
         .getSteps
         .get(0)
         .getProcessors
-        .get(1)
+        .get(0)
         .getOutputRows == 4)
 
     assert(
@@ -356,7 +355,7 @@ class GlutenClickHouseTPCHMetricsSuite extends GlutenClickHouseTPCHAbstractSuite
         .getSteps
         .get(0)
         .getName
-        .equals("ScalableMergingAggregated"))
+        .equals("GraceMergingAggregatedStep"))
     assert(
       nativeMetricsDataFinal.metricsDataList.get(1).getSteps.get(1).getName.equals("Expression"))
     assert(nativeMetricsDataFinal.metricsDataList.get(2).getName.equals("kProject"))

diff --git a/cpp-ch/local-engine/Operator/GraceMergingAggregatedStep.cpp b/cpp-ch/local-engine/Operator/GraceMergingAggregatedStep.cpp
@@ -58,6 +58,7 @@ GraceMergingAggregatedStep::GraceMergingAggregatedStep(
 
 void GraceMergingAggregatedStep::transformPipeline(DB::QueryPipelineBuilder & pipeline, const DB::BuildQueryPipelineSettings &)
 {
+    auto num_streams = pipeline.getNumStreams();
     auto transform_params = std::make_shared<DB::AggregatingTransformParams>(pipeline.getHeader(), params, true);
     pipeline.resize(1);
     auto build_transform = [&](DB::OutputPortRawPtrs outputs)
@@ -72,6 +73,7 @@ void GraceMergingAggregatedStep::transformPipeline(DB::QueryPipelineBuilder & pi
         return new_processors;
     };
     pipeline.transform(build_transform);
+    pipeline.resize(num_streams, true);
 }
 
 void GraceMergingAggregatedStep::describeActions(DB::IQueryPlanStep::FormatSettings & settings) const
@@ -225,6 +227,17 @@ void GraceMergingAggregatedTransform::extendBuckets()
 void GraceMergingAggregatedTransform::rehashDataVariants()
 {
     auto blocks = params->aggregator.convertToBlocks(*current_data_variants, false, 1);
+
+    size_t block_rows = 0;
+    size_t block_memory_usage = 0;
+    for (const auto & block : blocks)
+    {
+        block_rows += block.rows();
+        block_memory_usage += block.allocatedBytes();
+    }
+    if (block_rows)
+        per_key_memory_usage = block_memory_usage * 1.0 / block_rows;
+
     current_data_variants = std::make_shared<DB::AggregatedDataVariants>();
     no_more_keys = false;
     for (auto & block : blocks)
@@ -389,23 +402,48 @@ void GraceMergingAggregatedTransform::mergeOneBlock(const DB::Block &block)
 
 size_t GraceMergingAggregatedTransform::getMemoryUsage()
 {
-    size_t current_memory_usage = 0;
+    Int64 current_memory_usage = 0;
     if (auto * memory_tracker_child = DB::CurrentThread::getMemoryTracker())
         if (auto * memory_tracker = memory_tracker_child->getParent())
             current_memory_usage = memory_tracker->get();
-    return current_memory_usage;
+    return current_memory_usage < 0 ? 0 : current_memory_usage;
 }
 
 bool GraceMergingAggregatedTransform::isMemoryOverFlow()
 {
-    auto mem_limit = context->getSettingsRef().max_bytes_before_external_group_by;
-    if (!mem_limit)
+    /// More greedy memory usage strategy.
+    if (!context->getSettingsRef().max_memory_usage)
         return false;
-    auto current_memory_usage = getMemoryUsage();
-    if (current_memory_usage > mem_limit)
+    auto max_mem_used = context->getSettingsRef().max_memory_usage * 8 / 10;
+    auto current_result_rows = current_data_variants->size();
+    auto current_mem_used = getMemoryUsage();
+    if (per_key_memory_usage > 0)
     {
-        LOG_INFO(logger, "memory is overflow. used: {}, limit: {}", ReadableSize(current_memory_usage), ReadableSize(mem_limit));
-        return true;
+        if (current_mem_used + per_key_memory_usage * current_result_rows >= max_mem_used)
+        {
+            LOG_INFO(
+                logger,
+                "Memory is overflow. current_mem_used: {}, max_mem_used: {}, per_key_memory_usage: {}, aggregator keys: {}, buckets: {}",
+                ReadableSize(current_mem_used),
+                ReadableSize(max_mem_used),
+                ReadableSize(per_key_memory_usage),
+                current_result_rows,
+                getBucketsNum());
+            return true;
+        }
+    }
+    else
+    {
+        if (current_mem_used * 2 >= context->getSettingsRef().max_memory_usage)
+        {
+            LOG_INFO(
+                logger,
+                "Memory is overflow on half of max usage. current_mem_used: {}, max_mem_used: {}, buckets: {}",
+                ReadableSize(current_mem_used),
+                ReadableSize(context->getSettingsRef().max_memory_usage),
+                getBucketsNum());
+            return true;
+        }
     }
     return false;
 }

diff --git a/cpp-ch/local-engine/Operator/GraceMergingAggregatedStep.h b/cpp-ch/local-engine/Operator/GraceMergingAggregatedStep.h
@@ -97,6 +97,8 @@ class GraceMergingAggregatedTransform : public DB::IProcessor
     DB::BlocksList current_final_blocks;
     bool no_more_keys = false;
 
+    double per_key_memory_usage = 0;
+
     // metrics
     size_t total_input_blocks = 0;
     size_t total_input_rows = 0;