add session properties for prefixsort (prestodb#24043)

* add native session properties for prefixsort
unix280 · Nov 15, 2024 · e1bddb3 · e1bddb3
1 parent aee1a78
commit e1bddb3
Show file tree

Hide file tree

Showing 4 changed files with 96 additions and 0 deletions.
diff --git a/presto-docs/src/main/sphinx/presto_cpp/properties-session.rst b/presto-docs/src/main/sphinx/presto_cpp/properties-session.rst
@@ -276,3 +276,32 @@ bytes / number of destinations for each destination before producing a Serialize
 
 Maximum number of partitions created by a local exchange.
 Affects concurrency for pipelines containing LocalPartitionNode.
+
+
+``native_spill_prefixsort_enabled``
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+* **Type:** ``boolean``
+* **Default value:** ``false``
+
+Enable the prefix sort or fallback to std::sort in spill. The prefix sort is
+faster than std::sort but requires the memory to build normalized prefix
+keys, which might have potential risk of running out of server memory.
+
+``native_prefixsort_normalized_key_max_bytes``
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+* **Type:** ``integer``
+* **Default value:** ``128``
+
+Maximum number of bytes to use for the normalized key in prefix-sort.
+Use ``0`` to disable prefix-sort.
+
+``native_prefixsort_min_rows``
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+* **Type:** ``integer``
+* **Default value:** ``130``
+
+Minimum number of rows to use prefix-sort.
+The default value has been derived using micro-benchmarking.
diff --git a/...ava/com/facebook/presto/sessionpropertyproviders/NativeWorkerSessionPropertyProvider.java b/...ava/com/facebook/presto/sessionpropertyproviders/NativeWorkerSessionPropertyProvider.java
@@ -62,6 +62,9 @@ public class NativeWorkerSessionPropertyProvider
     public static final String NATIVE_QUERY_TRACE_MAX_BYTES = "native_query_trace_max_bytes";
     public static final String NATIVE_QUERY_TRACE_REG_EXP = "native_query_trace_task_reg_exp";
     public static final String NATIVE_MAX_LOCAL_EXCHANGE_PARTITION_COUNT = "native_max_local_exchange_partition_count";
+    public static final String NATIVE_SPILL_PREFIXSORT_ENABLED = "native_spill_prefixsort_enabled";
+    public static final String NATIVE_PREFIXSORT_NORMALIZED_KEY_MAX_BYTES = "native_prefixsort_normalized_key_max_bytes";
+    public static final String NATIVE_PREFIXSORT_MIN_ROWS = "native_prefixsort_min_rows";
     private final List<PropertyMetadata<?>> sessionProperties;
 
     @Inject
@@ -239,6 +242,25 @@ public NativeWorkerSessionPropertyProvider(FeaturesConfig featuresConfig)
                         "Maximum number of partitions created by a local exchange. " +
                                 "Affects concurrency for pipelines containing LocalPartitionNode",
                         null,
+                        !nativeExecution),
+                booleanProperty(
+                        NATIVE_SPILL_PREFIXSORT_ENABLED,
+                        "Enable the prefix sort or fallback to std::sort in spill. " +
+                                "The prefix sort is faster than std::sort but requires the memory to build normalized " +
+                                "prefix keys, which might have potential risk of running out of server memory.",
+                        false,
+                        !nativeExecution),
+                integerProperty(
+                        NATIVE_PREFIXSORT_NORMALIZED_KEY_MAX_BYTES,
+                        "Maximum number of bytes to use for the normalized key in prefix-sort. " +
+                                "Use 0 to disable prefix-sort.",
+                        128,
+                        !nativeExecution),
+                integerProperty(
+                        NATIVE_PREFIXSORT_MIN_ROWS,
+                        "Minimum number of rows to use prefix-sort. " +
+                                "The default value (130) has been derived using micro-benchmarking.",
+                        130,
                         !nativeExecution));
     }
 

diff --git a/presto-native-execution/presto_cpp/main/SessionProperties.cpp b/presto-native-execution/presto_cpp/main/SessionProperties.cpp
@@ -363,6 +363,34 @@ SessionProperties::SessionProperties() {
       false,
       QueryConfig::kMaxLocalExchangePartitionCount,
       std::to_string(c.maxLocalExchangePartitionCount()));
+
+  addSessionProperty(
+      kSpillPrefixSortEnabled,
+      "Enable the prefix sort or fallback to std::sort in spill. The prefix sort is "
+      "faster than std::sort but requires the memory to build normalized prefix "
+      "keys, which might have potential risk of running out of server memory.",
+      BOOLEAN(),
+      false,
+      QueryConfig::kSpillPrefixSortEnabled,
+      std::to_string(c.spillPrefixSortEnabled()));
+
+  addSessionProperty(
+      kPrefixSortNormalizedKeyMaxBytes,
+      "Maximum number of bytes to use for the normalized key in prefix-sort. "
+      "Use 0 to disable prefix-sort.",
+      INTEGER(),
+      false,
+      QueryConfig::kPrefixSortNormalizedKeyMaxBytes,
+      std::to_string(c.prefixSortNormalizedKeyMaxBytes()));
+
+  addSessionProperty(
+      kPrefixSortMinRows,
+      "Minimum number of rows to use prefix-sort. The default value (130) has been "
+      "derived using micro-benchmarking.",
+      INTEGER(),
+      false,
+      QueryConfig::kPrefixSortMinRows,
+      std::to_string(c.prefixSortMinRows()));
 }
 
 const std::unordered_map<std::string, std::shared_ptr<SessionProperty>>&

diff --git a/presto-native-execution/presto_cpp/main/SessionProperties.h b/presto-native-execution/presto_cpp/main/SessionProperties.h
@@ -225,6 +225,23 @@ class SessionProperties {
   static constexpr const char* kMaxLocalExchangePartitionCount =
       "native_max_local_exchange_partition_count";
 
+  /// Enable the prefix sort or fallback to std::sort in spill. The prefix sort
+  /// is faster than std::sort but requires the memory to build normalized
+  /// prefix keys, which might have potential risk of running out of server
+  /// memory.
+  static constexpr const char* kSpillPrefixSortEnabled =
+      "spill_prefixsort_enabled";
+
+  /// Maximum number of bytes to use for the normalized key in prefix-sort. Use
+  /// 0 to disable prefix-sort.
+  static constexpr const char* kPrefixSortNormalizedKeyMaxBytes =
+      "native_prefixsort_normalized_key_max_bytes";
+
+  /// Minimum number of rows to use prefix-sort. The default value (130) has
+  /// been derived using micro-benchmarking.
+  static constexpr const char* kPrefixSortMinRows =
+      "native_prefixsort_min_rows";
+
   SessionProperties();
 
   const std::unordered_map<std::string, std::shared_ptr<SessionProperty>>&