Merge pull request #21112 from vespa-engine/balder/use-mmap-for-large-vectors-3

vector of string_id tends to become very large. Use mmap allocation a…

11 files changed, 46 insertions, 38 deletions

diff --git a/eval/src/vespa/eval/eval/fast_addr_map.h b/eval/src/vespa/eval/eval/fast_addr_map.h
index a3b73afba6d..b9bc39ad619 100644
--- a/eval/src/vespa/eval/eval/fast_addr_map.h
+++ b/eval/src/vespa/eval/eval/fast_addr_map.h
@@ -7,7 +7,6 @@
 #include <vespa/vespalib/util/string_id.h>
 #include <vespa/vespalib/stllike/identity.h>
 #include <vespa/vespalib/stllike/hashtable.h>
-#include <vector>
 
 namespace vespalib::eval {
 
@@ -62,8 +61,8 @@ public:
     // view able to convert tags into sparse addresses
     struct LabelView {
         size_t addr_size;
-        const std::vector<string_id> &labels;
-        LabelView(size_t num_mapped_dims, const std::vector<string_id> &labels_in)
+        const StringIdVector &labels;
+        LabelView(size_t num_mapped_dims, const StringIdVector &labels_in)
             : addr_size(num_mapped_dims), labels(labels_in) {}
         ConstArrayRef<string_id> get_addr(size_t idx) const {
             return {&labels[idx * addr_size], addr_size};
@@ -105,7 +104,7 @@ private:
     HashType _map;
 
 public:
-    FastAddrMap(size_t num_mapped_dims, const std::vector<string_id> &labels_in, size_t expected_subspaces)
+    FastAddrMap(size_t num_mapped_dims, const StringIdVector &labels_in, size_t expected_subspaces)
         : _labels(num_mapped_dims, labels_in),
           _map(expected_subspaces * 2, Hash(), Equal(_labels)) {}
     ~FastAddrMap();
@@ -117,7 +116,7 @@ public:
     ConstArrayRef<string_id> get_addr(size_t idx) const { return _labels.get_addr(idx); }
     size_t size() const { return _map.size(); }
     constexpr size_t addr_size() const { return _labels.addr_size; }
-    const std::vector<string_id> &labels() const { return _labels.labels; }
+    const StringIdVector &labels() const { return _labels.labels; }
     template <typename T>
     size_t lookup(ConstArrayRef<T> addr, uint32_t hash) const {
         // assert(addr_size() == addr.size());
diff --git a/eval/src/vespa/eval/eval/fast_value.hpp b/eval/src/vespa/eval/eval/fast_value.hpp
index 185529b2f51..591350347e1 100644
--- a/eval/src/vespa/eval/eval/fast_value.hpp
+++ b/eval/src/vespa/eval/eval/fast_value.hpp
@@ -139,7 +139,7 @@ struct FastIterateView : public Value::Index::View {
 // operations by calling inline functions directly.
 struct FastValueIndex final : Value::Index {
     FastAddrMap map;
-    FastValueIndex(size_t num_mapped_dims_in, const std::vector<string_id> &labels, size_t expected_subspaces_in)
+    FastValueIndex(size_t num_mapped_dims_in, const StringIdVector &labels, size_t expected_subspaces_in)
         : map(num_mapped_dims_in, labels, expected_subspaces_in) {}
     size_t size() const override { return map.size(); }
     std::unique_ptr<View> create_view(ConstArrayRef<size_t> dims) const override;
@@ -213,13 +213,12 @@ struct FastCells {
 
 template <typename T, bool transient>
 struct FastValue final : Value, ValueBuilder<T> {
-
     using Handles = typename std::conditional<transient,
-                                     std::vector<string_id>,
+                                     StringIdVector,
                                      SharedStringRepo::Handles>::type;
 
-    static const std::vector<string_id> &get_view(const std::vector<string_id> &handles) { return handles; }
-    static const std::vector<string_id> &get_view(const SharedStringRepo::Handles &handles) { return handles.view(); }
+    static const StringIdVector &get_view(const StringIdVector &handles) { return handles; }
+    static const StringIdVector &get_view(const SharedStringRepo::Handles &handles) { return handles.view(); }
 
     ValueType my_type;
     size_t my_subspace_size;
@@ -227,14 +226,7 @@ struct FastValue final : Value, ValueBuilder<T> {
     FastValueIndex my_index;
     FastCells<T> my_cells;
 
-    FastValue(const ValueType &type_in, size_t num_mapped_dims_in, size_t subspace_size_in, size_t expected_subspaces_in)
-        : my_type(type_in), my_subspace_size(subspace_size_in),
-          my_handles(),
-          my_index(num_mapped_dims_in, get_view(my_handles), expected_subspaces_in),
-          my_cells(subspace_size_in * expected_subspaces_in)
-    {
-        my_handles.reserve(expected_subspaces_in * num_mapped_dims_in);
-    }
+    FastValue(const ValueType &type_in, size_t num_mapped_dims_in, size_t subspace_size_in, size_t expected_subspaces_in);
     ~FastValue() override;
     const ValueType &type() const override { return my_type; }
     const Value::Index &index() const override { return my_index; }
@@ -310,7 +302,20 @@ struct FastValue final : Value, ValueBuilder<T> {
         return usage;
     }
 };
-template <typename T,bool transient> FastValue<T,transient>::~FastValue() = default;
+
+template <typename T,bool transient>
+FastValue<T,transient>::FastValue(const ValueType &type_in, size_t num_mapped_dims_in,
+                                  size_t subspace_size_in, size_t expected_subspaces_in)
+    : my_type(type_in), my_subspace_size(subspace_size_in),
+      my_handles(),
+      my_index(num_mapped_dims_in, get_view(my_handles), expected_subspaces_in),
+      my_cells(subspace_size_in * expected_subspaces_in)
+{
+    my_handles.reserve(expected_subspaces_in * num_mapped_dims_in);
+}
+
+template <typename T,bool transient>
+FastValue<T,transient>::~FastValue() = default;
 
 //-----------------------------------------------------------------------------
 
diff --git a/eval/src/vespa/eval/eval/memory_usage_stuff.h b/eval/src/vespa/eval/eval/memory_usage_stuff.h
index 79a4cfb0eda..10f5459b0e7 100644
--- a/eval/src/vespa/eval/eval/memory_usage_stuff.h
+++ b/eval/src/vespa/eval/eval/memory_usage_stuff.h
@@ -10,8 +10,9 @@ namespace vespalib::eval {
 template <typename T>
 MemoryUsage self_memory_usage() { return MemoryUsage(sizeof(T), sizeof(T), 0, 0); }
 
-template <typename T>
-MemoryUsage vector_extra_memory_usage(const std::vector<T> &vec) {
+template <typename V>
+MemoryUsage vector_extra_memory_usage(const V &vec) {
+    using T = typename V::value_type;
     MemoryUsage usage;
     usage.incAllocatedBytes(sizeof(T) * vec.capacity());
     usage.incUsedBytes(sizeof(T) * vec.size());
diff --git a/eval/src/vespa/eval/streamed/streamed_value.h b/eval/src/vespa/eval/streamed/streamed_value.h
index ef44eeafce3..a1927e793b2 100644
--- a/eval/src/vespa/eval/streamed/streamed_value.h
+++ b/eval/src/vespa/eval/streamed/streamed_value.h
@@ -2,10 +2,10 @@
 
 #pragma once
 
+#include "streamed_value_index.h"
 #include <vespa/eval/eval/value_type.h>
 #include <vespa/eval/eval/value.h>
 #include <vespa/vespalib/util/shared_string_repo.h>
-#include "streamed_value_index.h"
 #include <cassert>
 
 namespace vespalib::eval {
diff --git a/eval/src/vespa/eval/streamed/streamed_value_index.h b/eval/src/vespa/eval/streamed/streamed_value_index.h
index 724874af577..d3f0f2781c4 100644
--- a/eval/src/vespa/eval/streamed/streamed_value_index.h
+++ b/eval/src/vespa/eval/streamed/streamed_value_index.h
@@ -16,10 +16,10 @@ class StreamedValueIndex : public Value::Index
 private:
     uint32_t _num_mapped_dims;
     uint32_t _num_subspaces;
-    const std::vector<string_id> &_labels_ref;
+    const StringIdVector &_labels_ref;
 
 public:
-    StreamedValueIndex(uint32_t num_mapped_dims, uint32_t num_subspaces, const std::vector<string_id> &labels_ref)
+    StreamedValueIndex(uint32_t num_mapped_dims, uint32_t num_subspaces, const StringIdVector &labels_ref)
         : _num_mapped_dims(num_mapped_dims),
           _num_subspaces(num_subspaces),
           _labels_ref(labels_ref)
diff --git a/eval/src/vespa/eval/streamed/streamed_value_utils.h b/eval/src/vespa/eval/streamed/streamed_value_utils.h
index efcaf4c6e7a..b808ad3c573 100644
--- a/eval/src/vespa/eval/streamed/streamed_value_utils.h
+++ b/eval/src/vespa/eval/streamed/streamed_value_utils.h
@@ -13,9 +13,9 @@ namespace vespalib::eval {
  *  Reading more labels than available will trigger an assert.
  **/
 struct LabelStream {
-    const std::vector<string_id> &source;
+    const StringIdVector &source;
     size_t pos;
-    LabelStream(const std::vector<string_id> &data) : source(data), pos(0) {}
+    LabelStream(const StringIdVector &data) : source(data), pos(0) {}
     string_id next_label() {
         assert(pos < source.size());
         return source[pos++];
@@ -42,7 +42,7 @@ private:
     size_t _num_subspaces;
     LabelStream _labels;
     size_t _subspace_index;
-    std::vector<string_id> _current_address;
+    StringIdVector _current_address;
 public:
     LabelBlock next_block() {
         if (_subspace_index < _num_subspaces) {
@@ -61,7 +61,7 @@ public:
     }
 
     LabelBlockStream(uint32_t num_subspaces,
-                     const std::vector<string_id> &labels,
+                     const StringIdVector &labels,
                      uint32_t num_mapped_dims)
       : _num_subspaces(num_subspaces),
         _labels(labels),
diff --git a/eval/src/vespa/eval/streamed/streamed_value_view.h b/eval/src/vespa/eval/streamed/streamed_value_view.h
index 908176d5cac..53b9e733de5 100644
--- a/eval/src/vespa/eval/streamed/streamed_value_view.h
+++ b/eval/src/vespa/eval/streamed/streamed_value_view.h
@@ -24,7 +24,7 @@ private:
 public:
     StreamedValueView(const ValueType &type, size_t num_mapped_dimensions,
                       TypedCells cells, size_t num_subspaces,
-                      const std::vector<string_id> &labels)
+                      const StringIdVector &labels)
       : _type(type),
         _cells_ref(cells),
         _my_index(num_mapped_dimensions, num_subspaces, labels)
diff --git a/searchlib/src/vespa/searchlib/tensor/streamed_value_store.cpp b/searchlib/src/vespa/searchlib/tensor/streamed_value_store.cpp
index 0901e643afa..2e6d771a870 100644
--- a/searchlib/src/vespa/searchlib/tensor/streamed_value_store.cpp
+++ b/searchlib/src/vespa/searchlib/tensor/streamed_value_store.cpp
@@ -11,7 +11,6 @@
 #include <vespa/vespalib/objects/nbostream.h>
 #include <vespa/vespalib/util/size_literals.h>
 #include <vespa/vespalib/util/typify.h>
-#include <vespa/vespalib/util/stringfmt.h>
 #include <vespa/log/log.h>
 
 LOG_SETUP(".searchlib.tensor.streamed_value_store");
@@ -22,6 +21,7 @@ using namespace vespalib::eval;
 using vespalib::ConstArrayRef;
 using vespalib::MemoryUsage;
 using vespalib::string_id;
+using vespalib::StringIdVector;
 
 namespace search::tensor {
 
@@ -61,12 +61,12 @@ struct MyFastValueView final : Value {
     const ValueType &my_type;
     FastValueIndex my_index;
     TypedCells my_cells;
-    MyFastValueView(const ValueType &type_ref, const std::vector<string_id> &handle_view, TypedCells cells, size_t num_mapped, size_t num_spaces)
+    MyFastValueView(const ValueType &type_ref, const StringIdVector &handle_view, TypedCells cells, size_t num_mapped, size_t num_spaces)
         : my_type(type_ref),
           my_index(num_mapped, handle_view, num_spaces),
           my_cells(cells)
     {
-        const std::vector<string_id> &labels = handle_view;
+        const StringIdVector &labels = handle_view;
         for (size_t i = 0; i < num_spaces; ++i) {
             ConstArrayRef<string_id> addr(&labels[i * num_mapped], num_mapped);
             my_index.map.add_mapping(FastAddrMap::hash_labels(addr));
diff --git a/vespalib/src/tests/shared_string_repo/shared_string_repo_test.cpp b/vespalib/src/tests/shared_string_repo/shared_string_repo_test.cpp
index 81c8271f755..5b267c3b9e9 100644
--- a/vespalib/src/tests/shared_string_repo/shared_string_repo_test.cpp
+++ b/vespalib/src/tests/shared_string_repo/shared_string_repo_test.cpp
@@ -101,8 +101,8 @@ std::unique_ptr<Handles> copy_strong_handles(const Handles &handles) {
     return result;
 }
 
-std::unique_ptr<std::vector<string_id>> make_weak_handles(const Handles &handles) {
-    return std::make_unique<std::vector<string_id>>(handles.view());
+std::unique_ptr<StringIdVector> make_weak_handles(const Handles &handles) {
+    return std::make_unique<StringIdVector>(handles.view());
 }
 
 //-----------------------------------------------------------------------------
@@ -202,7 +202,7 @@ struct Fixture {
             std::vector<vespalib::string> get_direct_result;
             std::unique_ptr<Handles> strong;
             std::unique_ptr<Handles> strong_copy;
-            std::unique_ptr<std::vector<string_id>> weak;
+            std::unique_ptr<StringIdVector> weak;
             auto copy_strings_task = [&](){ copy_strings_result = copy_strings(work); };
             auto copy_and_hash_task = [&](){ copy_and_hash_result = copy_and_hash(work); };
             auto local_enum_task = [&](){ local_enum_result = local_enum(work); };
diff --git a/vespalib/src/vespa/vespalib/util/shared_string_repo.h b/vespalib/src/vespa/vespalib/util/shared_string_repo.h
index ec65b942d88..d5faa0dfb0d 100644
--- a/vespalib/src/vespa/vespalib/util/shared_string_repo.h
+++ b/vespalib/src/vespa/vespalib/util/shared_string_repo.h
@@ -291,7 +291,7 @@ public:
     // A collection of string handles with ownership
     class Handles {
     private:
-        std::vector<string_id> _handles;
+        StringIdVector _handles;
     public:
         Handles();
         Handles(Handles &&rhs);
@@ -309,7 +309,7 @@ public:
             string_id id = _repo.copy(handle);
             _handles.push_back(id);
         }
-        const std::vector<string_id> &view() const { return _handles; }
+        const StringIdVector &view() const { return _handles; }
     };
 };
 
diff --git a/vespalib/src/vespa/vespalib/util/string_id.h b/vespalib/src/vespa/vespalib/util/string_id.h
index 7a72feee64a..7fec1da0bb8 100644
--- a/vespalib/src/vespa/vespalib/util/string_id.h
+++ b/vespalib/src/vespa/vespalib/util/string_id.h
@@ -2,7 +2,8 @@
 
 #pragma once
 
-#include <cstdint>
+#include <vespa/vespalib/stllike/allocator.h>
+#include <vector>
 
 namespace vespalib {
 
@@ -38,4 +39,6 @@ public:
     constexpr bool operator!=(const string_id &rhs) const noexcept { return (_id != rhs._id); }
 };
 
+using StringIdVector = std::vector<string_id, vespalib::allocator_large<string_id>>;
+
 }