From 729a399f3d33cdb729531f2a465a944f3f5e1bd8 Mon Sep 17 00:00:00 2001
From: Kajari Ghosh <ghoshkaj@gmail.com>
Date: Thu, 5 Apr 2018 08:11:05 -0400
Subject: [PATCH] pass in a hashed key to the threadlocal cache

500 mb threadlocal 2 t
---
 include/engine/data_watchdog.hpp      |  6 ++-
 include/engine/datafacade_factory.hpp |  3 +-
 include/engine/unpacking_cache.hpp    | 70 ++++++++++++++++++++-------
 3 files changed, 59 insertions(+), 20 deletions(-)

diff --git a/include/engine/data_watchdog.hpp b/include/engine/data_watchdog.hpp
index 97359844af0..a655c4a19d5 100644
--- a/include/engine/data_watchdog.hpp
+++ b/include/engine/data_watchdog.hpp
@@ -60,7 +60,8 @@ class DataWatchdogImpl<AlgorithmT, datafacade::ContiguousInternalMemoryDataFacad
                 DataFacadeFactory<datafacade::ContiguousInternalMemoryDataFacade, AlgorithmT>(
                     std::make_shared<datafacade::SharedMemoryAllocator>(
                         std::vector<storage::SharedRegionRegister::ShmKey>{
-                            static_region.shm_key, updatable_region.shm_key}), static_region.timestamp);
+                            static_region.shm_key, updatable_region.shm_key}),
+                    static_region.timestamp);
         }
 
         watcher = std::thread(&DataWatchdogImpl::Run, this);
@@ -115,7 +116,8 @@ class DataWatchdogImpl<AlgorithmT, datafacade::ContiguousInternalMemoryDataFacad
                 DataFacadeFactory<datafacade::ContiguousInternalMemoryDataFacade, AlgorithmT>(
                     std::make_shared<datafacade::SharedMemoryAllocator>(
                         std::vector<storage::SharedRegionRegister::ShmKey>{
-                            static_region.shm_key, updatable_region.shm_key}), static_region.timestamp);
+                            static_region.shm_key, updatable_region.shm_key}),
+                    static_region.timestamp);
         }
 
         util::Log() << "DataWatchdog thread stopped";
diff --git a/include/engine/datafacade_factory.hpp b/include/engine/datafacade_factory.hpp
index 67352538d80..5b459f31c29 100644
--- a/include/engine/datafacade_factory.hpp
+++ b/include/engine/datafacade_factory.hpp
@@ -71,7 +71,8 @@ template <template <typename A> class FacadeT, typename AlgorithmT> class DataFa
             std::size_t index =
                 std::stoi(exclude_prefix.substr(index_begin + 1, exclude_prefix.size()));
             BOOST_ASSERT(index >= 0 && index < facades.size());
-            facades[index] = std::make_shared<const Facade>(allocator, metric_name, index, timestamp);
+            facades[index] =
+                std::make_shared<const Facade>(allocator, metric_name, index, timestamp);
         }
 
         for (const auto index : util::irange<std::size_t>(0, properties->class_names.size()))
diff --git a/include/engine/unpacking_cache.hpp b/include/engine/unpacking_cache.hpp
index f043f18eb6d..487588f18a8 100644
--- a/include/engine/unpacking_cache.hpp
+++ b/include/engine/unpacking_cache.hpp
@@ -2,6 +2,7 @@
 #define UNPACKING_CACHE_HPP
 
 #include <boost/optional/optional_io.hpp>
+#include <boost/thread.hpp>
 
 #include "../../third_party/compute_detail/lru_cache.hpp"
 #include "util/typedefs.hpp"
@@ -11,28 +12,60 @@ namespace osrm
 namespace engine
 {
 typedef unsigned char ExcludeIndex;
+typedef unsigned Timestamp;
+typedef std::tuple<NodeID, NodeID, ExcludeIndex> Key;
+typedef std::size_t HashedKey;
+
+struct HashKey
+{
+    std::size_t operator()(Key const &key) const noexcept
+    {
+        std::size_t h1 = std::hash<NodeID>{}(std::get<0>(key));
+        std::size_t h2 = std::hash<NodeID>{}(std::get<1>(key));
+        std::size_t h3 = std::hash<ExcludeIndex>{}(std::get<2>(key));
+
+        std::size_t seed = 0;
+        boost::hash_combine(seed, h1);
+        boost::hash_combine(seed, h2);
+        boost::hash_combine(seed, h3);
+
+        return seed;
+    }
+};
+
 class UnpackingCache
 {
   private:
-    boost::compute::detail::lru_cache<std::tuple<NodeID, NodeID, ExcludeIndex>, EdgeDuration>
-        m_cache;
+    boost::compute::detail::lru_cache<HashedKey, EdgeDuration> m_cache;
     unsigned m_current_data_timestamp = 0;
 
   public:
     // TO FIGURE OUT HOW MANY LINES TO INITIALIZE CACHE TO:
     // Assume max cache size is 500mb (see bottom of OP here:
     // https://github.com/Project-OSRM/osrm-backend/issues/4798#issue-288608332)
+
+    // LRU CACHE IMPLEMENTATION HAS THESE TWO STORAGE CONTAINERS
+    // map: n * tuple_hash + n * EdgeDuration
+    //    = n * std::size_t + n * std::int32_t
+    //    = n * 8 bytes + n * 4 bytes
+    //    = n * 12 bytes
+    // list: n * HashedKey
+    //     = n * std::size_t
+    //     = n * 8 bytes
+    // Total = n * 20 bytes
     // Total cache size: 500 mb = 500 * 1024 *1024 bytes = 524288000 bytes
-    // Assume unsigned char is 1 byte (my local machine this is the case):
-    // Current cache line = NodeID * 2 + unsigned char * 1 + EdgeDuration * 1
-    //                    = std::uint32_t * 2 + unsigned char * 1 + std::int32_t * 1
-    //                    = 4 bytes * 3 + 1 byte = 13 bytes
-    // Number of cache lines is 500 mb = 500 * 1024 *1024 bytes = 524288000 bytes / 13 = 40329846
-    // For threadlocal cache, Number of cache lines = max cache size / number of threads
-    //                                              (Assume that the number of threads is 16)
-    //                                              = 40329846 / 16 = 2520615
 
-    UnpackingCache(unsigned timestamp) : m_cache(2520615), m_current_data_timestamp(timestamp){};
+    // THREAD LOCAL STORAGE
+    // Number of lines we need  = 524288000 / 20 / number of threads = 26214400 / number of threads
+    // 16 threads: 26214400 / 16 = 1638400
+    // 8 threads: 26214400 / 8 = 3276800
+    // 4 threads: 26214400 / 4 = 6553600
+    // 2 threads: 26214400 / 2 = 13107200
+
+    // SHARED STORAGE CACHE
+    // Number of lines we need for shared storage cache = 524288000 / 20 = 26214400
+
+    UnpackingCache(unsigned timestamp) : m_cache(13107200), m_current_data_timestamp(timestamp){};
 
     UnpackingCache(std::size_t cache_size, unsigned timestamp)
         : m_cache(cache_size), m_current_data_timestamp(timestamp){};
@@ -46,19 +79,22 @@ class UnpackingCache
         }
     }
 
-    bool IsEdgeInCache(std::tuple<NodeID, NodeID, ExcludeIndex> edge)
+    bool IsEdgeInCache(Key edge)
     {
-        return m_cache.contains(edge);
+        HashedKey hashed_edge = HashKey{}(edge);
+        return m_cache.contains(hashed_edge);
     }
 
-    void AddEdge(std::tuple<NodeID, NodeID, ExcludeIndex> edge, EdgeDuration duration)
+    void AddEdge(Key edge, EdgeDuration duration)
     {
-        m_cache.insert(edge, duration);
+        HashedKey hashed_edge = HashKey{}(edge);
+        m_cache.insert(hashed_edge, duration);
     }
 
-    EdgeDuration GetDuration(std::tuple<NodeID, NodeID, ExcludeIndex> edge)
+    EdgeDuration GetDuration(Key edge)
     {
-        boost::optional<EdgeDuration> duration = m_cache.get(edge);
+        HashedKey hashed_edge = HashKey{}(edge);
+        boost::optional<EdgeDuration> duration = m_cache.get(hashed_edge);
         return duration ? *duration : MAXIMAL_EDGE_DURATION;
     }
 };