Add single thread allocation mode (experimental).

CMakeLists.txt

@@ -101,6 +101,10 @@ set(INITIAL_SEGMENT_SIZE "0" CACHE STRING
         "Set the initial segment size (use the internally defined value if 0 is specified)")
 
 # ---------- Experimental options ---------- #
+# This mode still uses multiple threads inside Metall.
+# However, applications must not use metall with multiple threads.
+option(SINGLE_THREAD_ALLOC "Optimize Metall kernel for single thread usage" OFF)
+
 option(USE_ANONYMOUS_NEW_MAP "Use the anonymous map when creating a new map region" OFF)
 set(UMAP_ROOT "" CACHE PATH "UMap installed root directory")
 
@@ -211,6 +215,11 @@ if (USE_ANONYMOUS_NEW_MAP)
     message(STATUS "Use the anonymous map for new map region")
 endif ()
 
+if (SINGLE_THREAD_ALLOC)
+    list(APPEND METALL_DEFS "METALL_SINGLE_THREAD_ALLOC")
+    message(STATUS "Optimize Metall kernel for single thread usage")
+endif ()
+
 # Requirements for GCC
 if (NOT RUN_BUILD_AND_TEST_WITH_CI)
     if (("${CMAKE_C_COMPILER_ID}" STREQUAL "GNU") OR ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU"))

include/metall/kernel/object_cache.hpp

@@ -1,4 +1,4 @@
-// Copyright 2019 Lawrence Livermore National Security, LLC and other Metall
+// Copyright 2023 Lawrence Livermore National Security, LLC and other Metall
 // Project Developers. See the top-level COPYRIGHT file for details.
 //
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
@@ -19,10 +19,13 @@
 #include <metall/kernel/object_cache_container.hpp>
 #include <metall/detail/proc.hpp>
 #include <metall/detail/hash.hpp>
-#define ENABLE_MUTEX_IN_METALL_OBJECT_CACHE 1
-#if ENABLE_MUTEX_IN_METALL_OBJECT_CACHE
+
+#ifndef METALL_SINGLE_THREAD_ALLOC
+#define METALL_ENABLE_MUTEX_IN_OBJECT_CACHE
+#ifdef METALL_ENABLE_MUTEX_IN_OBJECT_CACHE
 #include <metall/detail/mutex.hpp>
 #endif
+#endif
 
 namespace metall::kernel {
 
@@ -54,7 +57,12 @@ class object_cache {
   // Private types and static values
   // -------------------- //
 
-  static constexpr std::size_t k_num_cache_per_core = 4;
+  static constexpr std::size_t k_num_cache_per_core =
+#ifdef METALL_SINGLE_THREAD_ALLOC
+      1;
+#else
+      4;
+#endif
   static constexpr std::size_t k_cache_bin_size = 1ULL << 20ULL;
   static constexpr std::size_t k_max_cache_block_size =
       64;  // Add and remove caches by up to this size
@@ -69,7 +77,7 @@ class object_cache {
                              difference_type, bin_no_manager>;
   using cache_table_type = std::vector<single_cache_type>;
 
-#if ENABLE_MUTEX_IN_METALL_OBJECT_CACHE
+#ifdef METALL_ENABLE_MUTEX_IN_OBJECT_CACHE
   using mutex_type = mdtl::mutex;
   using lock_guard_type = mdtl::mutex_lock_guard;
 #endif
@@ -85,7 +93,7 @@ class object_cache {
   // -------------------- //
   object_cache()
       : m_cache_table(get_num_cores() * k_num_cache_per_core)
-#if ENABLE_MUTEX_IN_METALL_OBJECT_CACHE
+#ifdef METALL_ENABLE_MUTEX_IN_OBJECT_CACHE
         ,
         m_mutex(m_cache_table.size())
 #endif
@@ -112,7 +120,7 @@ class object_cache {
     if (bin_no > max_bin_no()) return -1;
 
     const auto cache_no = priv_comp_cache_no();
-#if ENABLE_MUTEX_IN_METALL_OBJECT_CACHE
+#ifdef METALL_ENABLE_MUTEX_IN_OBJECT_CACHE
     lock_guard_type guard(m_mutex[cache_no]);
 #endif
     if (m_cache_table[cache_no].empty(bin_no)) {
@@ -140,7 +148,7 @@ class object_cache {
     if (bin_no > max_bin_no()) return false;  // Error
 
     const auto cache_no = priv_comp_cache_no();
-#if ENABLE_MUTEX_IN_METALL_OBJECT_CACHE
+#ifdef METALL_ENABLE_MUTEX_IN_OBJECT_CACHE
     lock_guard_type guard(m_mutex[cache_no]);
 #endif
     m_cache_table[cache_no].push(bin_no, object_offset);
@@ -224,6 +232,9 @@ class object_cache {
   }
 
   std::size_t priv_comp_cache_no() const {
+#ifdef METALL_SINGLE_THREAD_ALLOC
+    return 0;
+#endif
 #if SUPPORT_GET_CPU_CORE_NO
     thread_local static const auto sub_cache_no =
         std::hash<std::thread::id>{}(std::this_thread::get_id()) %
@@ -241,6 +252,9 @@ class object_cache {
   /// \brief Get CPU core number.
   /// This function does not call the system call every time as it is slow.
   static std::size_t priv_get_core_no() {
+#ifdef METALL_SINGLE_THREAD_ALLOC
+        return 0;
+#endif
     thread_local static int cached_core_no = 0;
     thread_local static int cached_count = 0;
     if (cached_core_no == 0) {
@@ -251,14 +265,18 @@ class object_cache {
   }
 
   static std::size_t get_num_cores() {
+#ifdef METALL_SINGLE_THREAD_ALLOC
+    return 1;
+#else
     return std::thread::hardware_concurrency();
+#endif
   }
 
   // -------------------- //
   // Private fields
   // -------------------- //
   cache_table_type m_cache_table;
-#if ENABLE_MUTEX_IN_METALL_OBJECT_CACHE
+#ifdef METALL_ENABLE_MUTEX_IN_OBJECT_CACHE
   std::vector<mutex_type> m_mutex;
 #endif
 };

test/kernel/CMakeLists.txt

@@ -10,6 +10,9 @@ add_metall_test_executable(chunk_directory_test chunk_directory_test.cpp)
 
 add_metall_test_executable(manager_test manager_test.cpp)
 
+add_metall_test_executable(manager_test_single_thread manager_test.cpp)
+target_compile_definitions(manager_test_single_thread PRIVATE METALL_SINGLE_THREAD_ALLOC)
+
 add_metall_test_executable(snapshot_test snapshot_test.cpp)
 
 add_metall_test_executable(copy_file_test copy_file_test.cpp)