Parallelize generation of the edge-expanded-edges.

Author: danpat

CHANGELOG.md

@@ -13,6 +13,7 @@
       - Added conditional restriction support with `parse-conditional-restrictions=true|false` to osrm-extract. This option saves conditional turn restrictions to the .restrictions file for parsing by contract later. Added `parse-conditionals-from-now=utc time stamp` and `--time-zone-file=/path/to/file`  to osrm-contract
       - Command-line tools (osrm-extract, osrm-contract, osrm-routed, etc) now return error codes and legible error messages for common problem scenarios, rather than ugly C++ crashes
       - Speed up pre-processing by only running the Lua `node_function` for nodes that have tags.  Cuts OSM file parsing time in half.
+      - osrm-extract now performs generation of edge-expanded-edges using all available CPUs, which should make osrm-extract significantly faster on multi-CPU machines
     - Files
       - .osrm.nodes file was renamed to .nbg_nodes and .ebg_nodes was added
     - Guidance

include/extractor/edge_based_graph_factory.hpp

@@ -18,6 +18,7 @@
 #include "extractor/query_node.hpp"
 #include "extractor/restriction_map.hpp"
 
+#include "util/concurrent_id_map.hpp"
 #include "util/deallocating_vector.hpp"
 #include "util/guidance/bearing_class.hpp"
 #include "util/guidance/entry_class.hpp"
@@ -41,6 +42,9 @@
 
 #include <boost/filesystem/fstream.hpp>
 
+#include <tbb/concurrent_unordered_map.h>
+#include <tbb/concurrent_vector.h>
+
 namespace osrm
 {
 namespace extractor
@@ -167,9 +171,9 @@ class EdgeBasedGraphFactory
     std::size_t skipped_uturns_counter;
     std::size_t skipped_barrier_turns_counter;
 
-    std::unordered_map<util::guidance::BearingClass, BearingClassID> bearing_class_hash;
+    util::ConcurrentIDMap<util::guidance::BearingClass, BearingClassID> bearing_class_hash;
     std::vector<BearingClassID> bearing_class_by_node_based_node;
-    std::unordered_map<util::guidance::EntryClass, EntryClassID> entry_class_hash;
+    util::ConcurrentIDMap<util::guidance::EntryClass, EntryClassID> entry_class_hash;
 };
 } // namespace extractor
 } // namespace osrm

include/extractor/guidance/turn_lane_handler.hpp

@@ -65,12 +65,16 @@ const constexpr char *scenario_names[] = {"Simple",
 
 class TurnLaneHandler
 {
+    using UpgradableMutex = boost::interprocess::interprocess_upgradable_mutex;
+    using ScopedReaderLock = boost::interprocess::sharable_lock<UpgradableMutex>;
+    using ScopedWriterLock = boost::interprocess::scoped_lock<UpgradableMutex>;
+
   public:
     typedef std::vector<TurnLaneData> LaneDataVector;
 
     TurnLaneHandler(const util::NodeBasedDynamicGraph &node_based_graph,
-                    std::vector<std::uint32_t> &turn_lane_offsets,
-                    std::vector<TurnLaneType::Mask> &turn_lane_masks,
+                    const std::vector<std::uint32_t> &turn_lane_offsets,
+                    const std::vector<TurnLaneType::Mask> &turn_lane_masks,
                     LaneDescriptionMap &lane_description_map,
                     const TurnAnalysis &turn_analysis,
                     util::guidance::LaneDataIdMap &id_map);
@@ -86,8 +90,8 @@ class TurnLaneHandler
     // we need to be able to look at previous intersections to, in some cases, find the correct turn
     // lanes for a turn
     const util::NodeBasedDynamicGraph &node_based_graph;
-    std::vector<std::uint32_t> &turn_lane_offsets;
-    std::vector<TurnLaneType::Mask> &turn_lane_masks;
+    const std::vector<std::uint32_t> &turn_lane_offsets;
+    const std::vector<TurnLaneType::Mask> &turn_lane_masks;
     LaneDescriptionMap &lane_description_map;
     const TurnAnalysis &turn_analysis;
     util::guidance::LaneDataIdMap &id_map;

include/extractor/guidance/turn_lane_types.hpp

@@ -10,6 +10,7 @@
 
 #include <boost/functional/hash.hpp>
 
+#include "util/concurrent_id_map.hpp"
 #include "util/json_container.hpp"
 #include "util/typedefs.hpp"
 
@@ -93,9 +94,9 @@ struct TurnLaneDescription_hash
     }
 };
 
-typedef std::unordered_map<guidance::TurnLaneDescription,
-                           LaneDescriptionID,
-                           guidance::TurnLaneDescription_hash>
+typedef util::ConcurrentIDMap<guidance::TurnLaneDescription,
+                              LaneDescriptionID,
+                              guidance::TurnLaneDescription_hash>
     LaneDescriptionMap;
 
 } // guidance

include/extractor/turn_data_container.hpp

@@ -31,6 +31,15 @@ void write(storage::io::FileWriter &writer,
            const detail::TurnDataContainerImpl<Ownership> &turn_data);
 }
 
+struct TurnData
+{
+    extractor::guidance::TurnInstruction turn_instruction;
+    LaneDataID lane_data_id;
+    EntryClassID entry_class_id;
+    util::guidance::TurnBearing pre_turn_bearing;
+    util::guidance::TurnBearing post_turn_bearing;
+};
+
 namespace detail
 {
 template <storage::Ownership Ownership> class TurnDataContainerImpl
@@ -75,17 +84,20 @@ template <storage::Ownership Ownership> class TurnDataContainerImpl
 
     // Used by EdgeBasedGraphFactory to fill data structure
     template <typename = std::enable_if<Ownership == storage::Ownership::Container>>
-    void push_back(extractor::guidance::TurnInstruction turn_instruction,
-                   LaneDataID lane_data_id,
-                   EntryClassID entry_class_id,
-                   util::guidance::TurnBearing pre_turn_bearing,
-                   util::guidance::TurnBearing post_turn_bearing)
+    void push_back(const TurnData &data)
+    {
+        turn_instructions.push_back(data.turn_instruction);
+        lane_data_ids.push_back(data.lane_data_id);
+        entry_class_ids.push_back(data.entry_class_id);
+        pre_turn_bearings.push_back(data.pre_turn_bearing);
+        post_turn_bearings.push_back(data.post_turn_bearing);
+    }
+
+    template <typename = std::enable_if<Ownership == storage::Ownership::Container>>
+    void append(const std::vector<TurnData> &others)
     {
-        turn_instructions.push_back(turn_instruction);
-        lane_data_ids.push_back(lane_data_id);
-        entry_class_ids.push_back(entry_class_id);
-        pre_turn_bearings.push_back(pre_turn_bearing);
-        post_turn_bearings.push_back(post_turn_bearing);
+        std::for_each(
+            others.begin(), others.end(), [this](const TurnData &other) { push_back(other); });
     }
 
     friend void serialization::read<Ownership>(storage::io::FileReader &reader,

include/util/concurrent_id_map.hpp

@@ -0,0 +1,57 @@
+#ifndef CONCURRENT_ID_MAP_HPP
+#define CONCURRENT_ID_MAP_HPP
+
+#include <boost/interprocess/sync/interprocess_upgradable_mutex.hpp>
+#include <boost/interprocess/sync/scoped_lock.hpp>
+#include <boost/interprocess/sync/sharable_lock.hpp>
+
+#include <unordered_map>
+
+namespace osrm
+{
+namespace util
+{
+
+/**
+ * This is a special purpose map for caching incrementing IDs
+ */
+template <typename KeyType, typename ValueType, typename HashType = std::hash<KeyType>>
+struct ConcurrentIDMap
+{
+    static_assert(std::is_unsigned<ValueType>::value, "Only unsigned integer types are supported.");
+
+    using UpgradableMutex = boost::interprocess::interprocess_upgradable_mutex;
+    using ScopedReaderLock = boost::interprocess::sharable_lock<UpgradableMutex>;
+    using ScopedWriterLock = boost::interprocess::scoped_lock<UpgradableMutex>;
+
+    std::unordered_map<KeyType, ValueType, HashType> data;
+    mutable UpgradableMutex mutex;
+
+    const ValueType ConcurrentFindOrAdd(const KeyType &key)
+    {
+        {
+            ScopedReaderLock sentry{mutex};
+            const auto result = data.find(key);
+            if (result != data.end())
+            {
+                return result->second;
+            }
+        }
+        {
+            ScopedWriterLock sentry{mutex};
+            const auto result = data.find(key);
+            if (result != data.end())
+            {
+                return result->second;
+            }
+            const auto id = static_cast<ValueType>(data.size());
+            data[key] = id;
+            return id;
+        }
+    }
+};
+
+} // util
+} // osrm
+
+#endif // CONCURRENT_ID_MAP_HPP

include/util/guidance/turn_lanes.hpp

@@ -7,6 +7,7 @@
 #include <unordered_map>
 #include <vector>
 
+#include "util/concurrent_id_map.hpp"
 #include "util/typedefs.hpp"
 
 #include <boost/functional/hash.hpp>
@@ -97,7 +98,7 @@ class LaneTupleIdPair
     }
 };
 
-using LaneDataIdMap = std::unordered_map<LaneTupleIdPair, LaneDataID, boost::hash<LaneTupleIdPair>>;
+using LaneDataIdMap = ConcurrentIDMap<LaneTupleIdPair, LaneDataID, boost::hash<LaneTupleIdPair>>;
 
 } // namespace guidance
 } // namespace util