Improve performance of ASOF join

src/Common/RadixSort.h

@@ -515,6 +515,11 @@ public:
         radixSortLSDInternal<false>(arr, size, false, nullptr);
     }
 
+    static void executeLSD(Element * arr, size_t size, bool reverse)
+    {
+        radixSortLSDInternal<false>(arr, size, reverse, nullptr);
+    }
+
     /** This function will start to sort inplace (modify 'arr')
       *  but on the last step it will write result directly to the destination
       *  instead of finishing sorting 'arr'.

src/Interpreters/HashJoin.cpp

@@ -1237,16 +1237,16 @@ NO_INLINE IColumn::Filter joinRightColumns(
                 {
                     const IColumn & left_asof_key = added_columns.leftAsofKey();
 
-                    auto [block, row_num] = mapped->findAsof(left_asof_key, i);
-                    if (block)
+                    auto row_ref = mapped->findAsof(left_asof_key, i);
+                    if (row_ref.block)
                     {
                         setUsed<need_filter>(filter, i);
                         if constexpr (multiple_disjuncts)
-                            used_flags.template setUsed<jf.need_flags, multiple_disjuncts>(block, row_num, 0);
+                            used_flags.template setUsed<jf.need_flags, multiple_disjuncts>(row_ref.block, row_ref.row_num, 0);
                         else
                             used_flags.template setUsed<jf.need_flags, multiple_disjuncts>(find_result);
 
-                        added_columns.appendFromBlock<jf.add_missing>(*block, row_num);
+                        added_columns.appendFromBlock<jf.add_missing>(*row_ref.block, row_ref.row_num);
                     }
                     else
                         addNotFoundRow<jf.add_missing, jf.need_replication>(added_columns, current_offset);

src/Interpreters/RowRefs.cpp

@@ -1,5 +1,6 @@
 #include <Interpreters/RowRefs.h>
 
+#include <Common/RadixSort.h>
 #include <AggregateFunctions/Helpers.h>
 #include <Columns/IColumn.h>
 #include <DataTypes/IDataType.h>
@@ -44,38 +45,52 @@ class SortedLookupVector : public SortedLookupVectorBase
 {
     struct Entry
     {
-        /// We don't store a RowRef and instead keep it's members separately (and return a tuple) to reduce the memory usage.
-        /// For example, for sizeof(T) == 4 => sizeof(Entry) == 16 (while before it would be 20). Then when you put it into a vector, the effect is even greater
-        decltype(RowRef::block) block;
-        decltype(RowRef::row_num) row_num;
-        TKey asof_value;
+        TKey value;
+        uint32_t row_ref_index;
 
         Entry() = delete;
-        Entry(TKey v, const Block * b, size_t r) : block(b), row_num(r), asof_value(v) { }
+        Entry(TKey value_, uint32_t row_ref_index_)
+            : value(value_)
+            , row_ref_index(row_ref_index_)
+        { }
 
-        bool operator<(const Entry & other) const { return asof_value < other.asof_value; }
+    };
+
+    struct LessEntryOperator
+    {
+        ALWAYS_INLINE bool operator()(const Entry & lhs, const Entry & rhs) const
+        {
+            return lhs.value > rhs.value;
+        }
     };
 
     struct GreaterEntryOperator
     {
-        bool operator()(Entry const & a, Entry const & b) const { return a.asof_value > b.asof_value; }
+        ALWAYS_INLINE bool operator()(const Entry & lhs, const Entry & rhs) const
+        {
+            return lhs.value > rhs.value;
+        }
     };
 
 
 public:
-    using Base = std::vector<Entry>;
     using Keys = std::vector<TKey>;
-    static constexpr bool isDescending = (inequality == ASOF::Inequality::Greater || inequality == ASOF::Inequality::GreaterOrEquals);
-    static constexpr bool isStrict = (inequality == ASOF::Inequality::Less) || (inequality == ASOF::Inequality::Greater);
+    using Entries = PaddedPODArray<Entry>;
+    using RowRefs = PaddedPODArray<RowRef>;
+
+    static constexpr bool is_descending = (inequality == ASOF::Inequality::Greater || inequality == ASOF::Inequality::GreaterOrEquals);
+    static constexpr bool is_strict = (inequality == ASOF::Inequality::Less) || (inequality == ASOF::Inequality::Greater);
 
     void insert(const IColumn & asof_column, const Block * block, size_t row_num) override
     {
         using ColumnType = ColumnVectorOrDecimal<TKey>;
         const auto & column = assert_cast<const ColumnType &>(asof_column);
-        TKey k = column.getElement(row_num);
+        TKey key = column.getElement(row_num);
 
         assert(!sorted.load(std::memory_order_acquire));
-        array.emplace_back(k, block, row_num);
+
+        entries.emplace_back(key, row_refs.size());
+        row_refs.emplace_back(RowRef(block, row_num));
     }
 
     /// Unrolled version of upper_bound and lower_bound
@@ -84,30 +99,30 @@ public:
     /// at https://en.algorithmica.org/hpc/data-structures/s-tree/
     size_t boundSearch(TKey value)
     {
-        size_t size = array.size();
+        size_t size = entries.size();
         size_t low = 0;
 
         /// This is a single binary search iteration as a macro to unroll. Takes into account the inequality:
-        /// isStrict -> Equal values are not requested
-        /// isDescending -> The vector is sorted in reverse (for greater or greaterOrEquals)
+        /// is_strict -> Equal values are not requested
+        /// is_descending -> The vector is sorted in reverse (for greater or greaterOrEquals)
 #define BOUND_ITERATION \
     { \
         size_t half = size / 2; \
         size_t other_half = size - half; \
         size_t probe = low + half; \
         size_t other_low = low + other_half; \
-        TKey v = array[probe].asof_value; \
+        TKey & v = entries[probe].value; \
         size = half; \
-        if constexpr (isDescending) \
+        if constexpr (is_descending) \
         { \
-            if constexpr (isStrict) \
+            if constexpr (is_strict) \
                 low = value <= v ? other_low : low; \
             else \
                 low = value < v ? other_low : low; \
         } \
         else \
         { \
-            if constexpr (isStrict) \
+            if constexpr (is_strict) \
                 low = value >= v ? other_low : low; \
             else \
                 low = value > v ? other_low : low; \
@@ -130,7 +145,7 @@ public:
         return low;
     }
 
-    std::tuple<decltype(RowRef::block), decltype(RowRef::row_num)> findAsof(const IColumn & asof_column, size_t row_num) override
+    RowRef findAsof(const IColumn & asof_column, size_t row_num) override
     {
         sort();
 
@@ -139,8 +154,11 @@ public:
         TKey k = column.getElement(row_num);
 
         size_t pos = boundSearch(k);
-        if (pos != array.size())
-            return std::make_tuple(array[pos].block, array[pos].row_num);
+        if (pos != entries.size())
+        {
+            size_t row_ref_index = entries[pos].row_ref_index;
+            return row_refs[row_ref_index];
+        }
 
         return {nullptr, 0};
     }
@@ -148,7 +166,8 @@ public:
 private:
     std::atomic<bool> sorted = false;
     mutable std::mutex lock;
-    Base array;
+    Entries entries;
+    RowRefs row_refs;
 
     // Double checked locking with SC atomics works in C++
     // https://preshing.com/20130930/double-checked-locking-is-fixed-in-cpp11/
@@ -160,12 +179,37 @@ private:
         if (!sorted.load(std::memory_order_acquire))
         {
             std::lock_guard<std::mutex> l(lock);
+
             if (!sorted.load(std::memory_order_relaxed))
             {
-                if constexpr (isDescending)
-                    ::sort(array.begin(), array.end(), GreaterEntryOperator());
+                if constexpr (std::is_arithmetic_v<TKey> && !std::is_floating_point_v<TKey>)
+                {
+                    if (likely(entries.size() > 256))
+                    {
+                        struct RadixSortTraits : RadixSortNumTraits<TKey>
+                        {
+                            using Element = Entry;
+                            using Result = Element;
+
+                            static TKey & extractKey(Element & elem) { return elem.value; }
+                            static Element extractResult(Element & elem) { return elem; }
+                        };
+
+                        if constexpr (is_descending)
+                            RadixSort<RadixSortTraits>::executeLSD(entries.data(), entries.size(), true);
+                        else
+                            RadixSort<RadixSortTraits>::executeLSD(entries.data(), entries.size(), false);
+
+                        sorted.store(true, std::memory_order_release);
+                        return;
+                    }
+                }
+
+                if constexpr (is_descending)
+                    ::sort(entries.begin(), entries.end(), GreaterEntryOperator());
                 else
-                    ::sort(array.begin(), array.end());
+                    ::sort(entries.begin(), entries.end(), LessEntryOperator());
+
                 sorted.store(true, std::memory_order_release);
             }
         }

src/Interpreters/RowRefs.h

@@ -146,7 +146,7 @@ private:
 struct SortedLookupVectorBase
 {
     SortedLookupVectorBase() = default;
-    virtual ~SortedLookupVectorBase() { }
+    virtual ~SortedLookupVectorBase() = default;
 
     static std::optional<TypeIndex> getTypeSize(const IColumn & asof_column, size_t & type_size);
 
@@ -154,7 +154,7 @@ struct SortedLookupVectorBase
     virtual void insert(const IColumn &, const Block *, size_t) = 0;
 
     // This needs to be synchronized internally
-    virtual std::tuple<decltype(RowRef::block), decltype(RowRef::row_num)> findAsof(const IColumn &, size_t) = 0;
+    virtual RowRef findAsof(const IColumn &, size_t) = 0;
 };