optimize the memory allocator

include/pool_alloc.h

@@ -11,9 +11,8 @@ namespace mem {
 
 class IPC_EXPORT pool_alloc {
 public:
-    static void  clear();
     static void* alloc(std::size_t size);
-    static void  free(void* p, std::size_t size);
+    static void  free (void* p, std::size_t size);
 };
 
 ////////////////////////////////////////////////////////////////

src/memory/alloc.h

@@ -12,13 +12,14 @@
 
 #include "platform/detail.h"
 
+#include "log.h"
+
 namespace ipc {
 namespace mem {
 
 class static_alloc {
 public:
     static void swap(static_alloc&) {}
-    static void clear() {}
 
     static void* alloc(std::size_t size) {
         return size ? std::malloc(size) : nullptr;
@@ -104,16 +105,11 @@ private:
 public:
     scope_alloc() = default;
 
-    scope_alloc(scope_alloc&& rhs)          { swap(rhs); }
+    scope_alloc(scope_alloc && rhs)         { swap(rhs); }
     scope_alloc& operator=(scope_alloc rhs) { swap(rhs); return (*this); }
 
     ~scope_alloc() { free_all(); }
 
-    template <typename A>
-    void set_allocator(A && alc) {
-        alloc_ = std::forward<A>(alc);
-    }
-
     void swap(scope_alloc& rhs) {
         alloc_.swap(rhs.alloc_);
         base_t::swap(rhs);
@@ -130,12 +126,6 @@ public:
         base_t::take(std::move(rhs));
     }
 
-    void clear() {
-        free_all();
-        tail_ = nullptr;
-        alloc_.~alloc_policy();
-    }
-
     void* alloc(std::size_t size) {
         auto curr = static_cast<block_t*>(alloc_.alloc(size += aligned_block_size));
         curr->next_ = head_;
@@ -156,10 +146,12 @@ namespace detail {
 
 class fixed_alloc_base {
 protected:
+    std::size_t block_size_;
     std::size_t init_expand_;
     void      * cursor_;
 
-    void init(std::size_t init_expand) {
+    void init(std::size_t block_size, std::size_t init_expand) {
+        block_size_  = block_size;
         init_expand_ = init_expand;
         cursor_      = nullptr;
     }
@@ -173,7 +165,12 @@ protected:
     }
 
 public:
+    void set_block_size(std::size_t block_size) {
+        block_size_ = block_size;
+    }
+
     void swap(fixed_alloc_base& rhs) {
+        std::swap(block_size_ , rhs.block_size_);
         std::swap(init_expand_, rhs.init_expand_);
         std::swap(cursor_     , rhs.cursor_);
     }
@@ -211,6 +208,63 @@ public:
     }
 };
 
+template <typename AllocP, typename ExpandP>
+class fixed_alloc : public detail::fixed_alloc_base {
+public:
+    using base_t = detail::fixed_alloc_base;
+    using alloc_policy = AllocP;
+
+private:
+    alloc_policy alloc_;
+
+    void* try_expand() {
+        if (empty()) {
+            auto size = ExpandP::next(block_size_, init_expand_);
+            auto p = node_p(cursor_ = alloc_.alloc(size));
+            for (std::size_t i = 0; i < (size / block_size_) - 1; ++i)
+                p = node_p((*p) = reinterpret_cast<byte_t*>(p) + block_size_);
+            (*p) = nullptr;
+        }
+        return cursor_;
+    }
+
+public:
+    explicit fixed_alloc(std::size_t block_size, std::size_t init_expand = 1) {
+        init(block_size, init_expand);
+    }
+
+    fixed_alloc(fixed_alloc && rhs) {
+        init(0, 0);
+        swap(rhs);
+    }
+
+    fixed_alloc& operator=(fixed_alloc rhs) {
+        swap(rhs);
+        return (*this);
+    }
+
+    void swap(fixed_alloc& rhs) {
+        alloc_.swap(rhs.alloc_);
+        base_t::swap(rhs);
+    }
+
+    template <typename A = AllocP>
+    auto take(fixed_alloc && rhs) -> ipc::require<detail::has_take<A>::value> {
+        base_t::take(std::move(rhs));
+        alloc_.take(std::move(rhs.alloc_));
+    }
+
+    void* alloc() {
+        void* p = try_expand();
+        cursor_ = next(p);
+        return p;
+    }
+
+    void* alloc(std::size_t) {
+        return alloc();
+    }
+};
+
 } // namespace detail
 
 struct fixed_expand_policy {
@@ -227,9 +281,8 @@ struct fixed_expand_policy {
         return e * 2;
     }
 
-    template <std::size_t BlockSize>
+    static std::size_t next(std::size_t block_size, std::size_t & e) {
-    static std::size_t next(std::size_t & e) {
+        auto n = ipc::detail::max<std::size_t>(block_size, base_size) * e;
-        auto n = ipc::detail::max<std::size_t>(BlockSize, base_size) * e;
         e = next(e);
         return n;
     }
@@ -238,73 +291,31 @@ struct fixed_expand_policy {
 template <std::size_t BlockSize,
           typename AllocP  = scope_alloc<>,
           typename ExpandP = fixed_expand_policy>
-class fixed_alloc : public detail::fixed_alloc_base {
+class fixed_alloc : public detail::fixed_alloc<AllocP, ExpandP> {
 public:
-    using base_t = detail::fixed_alloc_base;
+    using base_t = detail::fixed_alloc<AllocP, ExpandP>;
-    using alloc_policy = AllocP;
 
     enum : std::size_t {
         block_size = (ipc::detail::max)(BlockSize, sizeof(void*))
     };
 
-private:
-    alloc_policy alloc_;
-
-    void* try_expand() {
-        if (empty()) {
-            auto size = ExpandP::template next<block_size>(init_expand_);
-            auto p = node_p(cursor_ = alloc_.alloc(size));
-            for (std::size_t i = 0; i < (size / block_size) - 1; ++i)
-                p = node_p((*p) = reinterpret_cast<byte_t*>(p) + block_size);
-            (*p) = nullptr;
-        }
-        return cursor_;
-    }
-
 public:
-    explicit fixed_alloc(std::size_t init_expand = 1) {
+    explicit fixed_alloc(std::size_t init_expand = 1)
-        init(init_expand);
+        : base_t(block_size) {
     }
 
-    fixed_alloc(fixed_alloc&& rhs) : fixed_alloc() { swap(rhs); }
+    fixed_alloc(fixed_alloc && rhs)
-    fixed_alloc& operator=(fixed_alloc rhs)        { swap(rhs); return (*this); }
+        : base_t(std::move(rhs)) {
+    }
 
-    template <typename A>
+    fixed_alloc& operator=(fixed_alloc rhs) {
-    void set_allocator(A && alc) {
+        swap(rhs);
-        alloc_ = std::forward<A>(alc);
+        return (*this);
     }
 
     void swap(fixed_alloc& rhs) {
-        alloc_.swap(rhs.alloc_);
         base_t::swap(rhs);
     }
-
-    template <typename A = AllocP>
-    auto take(fixed_alloc && rhs) -> ipc::require<detail::has_take<A>::value> {
-        base_t::take(std::move(rhs));
-        alloc_.take(std::move(rhs.alloc_));
-    }
-
-    template <typename A = AllocP>
-    auto take(fixed_alloc && rhs) -> ipc::require<!detail::has_take<A>::value> {
-        base_t::take(std::move(rhs));
-    }
-
-    void clear() {
-        init_expand_ = ExpandP::prev(init_expand_);
-        cursor_ = nullptr;
-        alloc_.~alloc_policy();
-    }
-
-    void* alloc() {
-        void* p = try_expand();
-        cursor_ = next(p);
-        return p;
-    }
-
-    void* alloc(std::size_t) {
-        return alloc();
-    }
 };
 
 ////////////////////////////////////////////////////////////////
@@ -374,14 +385,9 @@ private:
 public:
     variable_alloc() = default;
 
-    variable_alloc(variable_alloc&& rhs)          { swap(rhs); }
+    variable_alloc(variable_alloc && rhs)         { swap(rhs); }
     variable_alloc& operator=(variable_alloc rhs) { swap(rhs); return (*this); }
 
-    template <typename A>
-    void set_allocator(A && alc) {
-        alloc_ = std::forward<A>(alc);
-    }
-
     void swap(variable_alloc& rhs) {
         alloc_.swap(rhs.alloc_);
         base_t::swap(rhs);
@@ -398,10 +404,6 @@ public:
         base_t::take(std::move(rhs));
     }
 
-    void clear() {
-        alloc_.~alloc_policy();
-    }
-
     void* alloc(std::size_t size) {
         if (size >= ChunkSize) {
             return alloc_.alloc(size);

src/memory/resource.h

@@ -21,19 +21,9 @@ using chunk_variable_alloc =
       static_wrapper<async_wrapper<variable_alloc<
                                    sizeof(void*) * 1024 * 256 /* 2MB(x64) */ >>>;
 
-template <std::size_t Size>
+using async_pool_alloc = 
-using static_async_fixed =
+      static_variable_wrapper<async_wrapper<detail::fixed_alloc<
-      static_wrapper<async_wrapper<fixed_alloc<
+                                            chunk_variable_alloc, fixed_expand_policy>>>;
-                                   Size, chunk_variable_alloc >>>;
-
-using big_size_alloc   = variable_wrapper<static_async_fixed,
-                                          default_mapping_policy<
-                                          default_mapping_policy<>::block_size(default_mapping_policy<>::classes_size),
-                                          default_mapping_policy<>::iter_size * 2 >>;
-
-using async_pool_alloc = variable_wrapper<static_async_fixed,
-                                          default_mapping_policy<>,
-                                          big_size_alloc>;
 
 template <typename T>
 using allocator = allocator_wrapper<T, async_pool_alloc>;

src/memory/wrapper.h

@@ -1,14 +1,15 @@
 #pragma once
 
 #include <limits>
-#include <new>
+#include <new>          // ::new
 #include <tuple>
 #include <thread>
 #include <vector>
-#include <functional>
+#include <functional>   // std::function
-#include <utility>
+#include <utility>      // std::forward
 #include <cstddef>
-#include <type_traits>
+#include <cassert>      // assert
+#include <type_traits>  // std::aligned_storage_t
 
 #include "def.h"
 #include "rw_lock.h"
@@ -80,27 +81,27 @@ public:
     struct rebind { typedef allocator_wrapper<U, AllocP> other; };
 
     constexpr size_type max_size(void) const noexcept {
-        return (std::numeric_limits<size_type>::max)() / sizeof(T);
+        return (std::numeric_limits<size_type>::max)() / sizeof(value_type);
     }
 
 public:
     pointer allocate(size_type count) noexcept {
         if (count == 0) return nullptr;
         if (count > this->max_size()) return nullptr;
-        return static_cast<pointer>(alloc_.alloc(count * sizeof(T)));
+        return static_cast<pointer>(alloc_.alloc(count * sizeof(value_type)));
     }
 
     void deallocate(pointer p, size_type count) noexcept {
-        alloc_.free(p, count * sizeof(T));
+        alloc_.free(p, count * sizeof(value_type));
     }
 
     template <typename... P>
-    static void construct(pointer p, P&&... params) {
+    static void construct(pointer p, P && ... params) {
-        ::new (static_cast<void*>(p)) T(std::forward<P>(params)...);
+        ::new (static_cast<void*>(p)) value_type(std::forward<P>(params) ...);
     }
 
     static void destroy(pointer p) {
-        p->~T();
+        p->~value_type();
     }
 };
 
@@ -148,11 +149,6 @@ public:
         master_allocs_.swap(rhs.master_allocs_);
     }
 
-    void clear() {
-        IPC_UNUSED_ auto guard = ipc::detail::unique_lock(master_lock_);
-        master_allocs_.clear();
-    }
-
     void try_recover(alloc_policy & alc) {
         IPC_UNUSED_ auto guard = ipc::detail::unique_lock(master_lock_);
         if (!master_allocs_.empty()) {
@@ -174,18 +170,16 @@ public:
 
     template <typename A = AllocP>
     auto try_replenish(alloc_policy & alc, std::size_t /*size*/)
-        -> ipc::require<detail::has_take<A>::value && !has_remain<A>::value && has_empty<A>::value> {
+        -> ipc::require<(!detail::has_take<A>::value || !has_remain<A>::value) && has_empty<A>::value> {
         if (!alc.empty()) return;
-        IPC_UNUSED_ auto guard = ipc::detail::unique_lock(master_lock_);
+        try_recover(alc);
-        if (!master_allocs_.empty()) {
-            alc.take(std::move(master_allocs_.back()));
-            master_allocs_.pop_back();
-        }
     }
 
     template <typename A = AllocP>
     constexpr auto try_replenish(alloc_policy & /*alc*/, std::size_t /*size*/) const noexcept
-        -> ipc::require<!detail::has_take<A>::value || (!has_remain<A>::value && !has_empty<A>::value)> {}
+        -> ipc::require<(!detail::has_take<A>::value || !has_remain<A>::value) && !has_empty<A>::value> {
+        // Do Nothing.
+    }
 
     void collect(alloc_policy && alc) {
         IPC_UNUSED_ auto guard = ipc::detail::unique_lock(master_lock_);
@@ -199,7 +193,6 @@ public:
     using alloc_policy = AllocP;
 
     constexpr static void swap(empty_alloc_recycler&)               noexcept {}
-    constexpr static void clear()                                   noexcept {}
     constexpr static void try_recover(alloc_policy&)                noexcept {}
     constexpr static auto try_replenish(alloc_policy&, std::size_t) noexcept {}
     constexpr static void collect(alloc_policy&&)                   noexcept {}
@@ -218,51 +211,47 @@ private:
         async_wrapper * w_ = nullptr;
 
     public:
-        alloc_proxy(alloc_proxy && rhs)
+        alloc_proxy(alloc_proxy && rhs) = default;
-            : AllocP(std::move(rhs))
-        {}
 
-        alloc_proxy(async_wrapper* w)
+        template <typename ... P>
-            : AllocP(), w_(w) {
+        alloc_proxy(async_wrapper* w, P && ... pars)
-            if (w_ == nullptr) return;
+            : AllocP(std::forward<P>(pars) ...), w_(w) {
+            assert(w_ != nullptr);
             w_->recycler_.try_recover(*this);
         }
 
         ~alloc_proxy() {
-            if (w_ == nullptr) return;
             w_->recycler_.collect(std::move(*this));
         }
 
-        auto alloc(std::size_t size) {
+        // auto alloc(std::size_t size) {
-            if (w_ != nullptr) {
+        //     w_->recycler_.try_replenish(*this, size);
-                w_->recycler_.try_replenish(*this, size);
+        //     return AllocP::alloc(size);
-            }
+        // }
-            return AllocP::alloc(size);
-        }
     };
 
     friend class alloc_proxy;
 
-    auto& get_alloc() {
+    using ref_t = alloc_proxy&;
-        static tls::pointer<alloc_proxy> tls_alc;
+    using tls_t = tls::pointer<alloc_proxy>;
-        return *tls_alc.create(this);
+
-    }
+    tls_t tls_;
+    std::function<ref_t()> get_alloc_;
 
 public:
-    void swap(async_wrapper& rhs) {
+    template <typename ... P>
-        recycler_.swap(rhs.recycler_);
+    async_wrapper(P ... pars) {
-    }
+        get_alloc_ = [this, pars ...]()->ref_t {
-
+            return *tls_.create(this, pars ...);
-    void clear() {
+        };
-        recycler_.clear();
     }
 
     void* alloc(std::size_t size) {
-        return get_alloc().alloc(size);
+        return get_alloc_().alloc(size);
     }
 
     void free(void* p, std::size_t size) {
-        get_alloc().free(p, size);
+        get_alloc_().free(p, size);
     }
 };
 
@@ -281,16 +270,16 @@ private:
     alloc_policy alloc_;
 
 public:
+    template <typename ... P>
+    sync_wrapper(P && ... pars)
+        : alloc_(std::forward<P>(pars) ...)
+    {}
+
     void swap(sync_wrapper& rhs) {
         IPC_UNUSED_ auto guard = ipc::detail::unique_lock(lock_);
         alloc_.swap(rhs.alloc_);
     }
 
-    void clear() {
-        IPC_UNUSED_ auto guard = ipc::detail::unique_lock(lock_);
-        alloc_.~alloc_policy();
-    }
-
     void* alloc(std::size_t size) {
         IPC_UNUSED_ auto guard = ipc::detail::unique_lock(lock_);
         return alloc_.alloc(size);
@@ -318,10 +307,6 @@ public:
 
     static void swap(static_wrapper&) {}
 
-    static void clear() {
-        instance().clear();
-    }
-
     static void* alloc(std::size_t size) {
         return instance().alloc(size);
     }
@@ -341,62 +326,56 @@ struct default_mapping_policy {
     enum : std::size_t {
         base_size    = BaseSize,
         iter_size    = IterSize,
-        classes_size = 32
+        classes_size = 64
     };
 
-    static const std::size_t table[classes_size];
+    template <typename F, typename ... P>
-
+    IPC_CONSTEXPR_ static void foreach(F f, P ... params) {
-    IPC_CONSTEXPR_ static std::size_t classify(std::size_t size) noexcept {
+        for (std::size_t i = 0; i < classes_size; ++i) f(i, params...);
-        auto index = (size <= base_size) ? 0 : ((size - base_size - 1) / iter_size);
-        return (index < classes_size) ?
-            // always uses default_mapping_policy<>::table
-            default_mapping_policy<>::table[index] : classes_size;
     }
 
-    constexpr static std::size_t block_size(std::size_t value) noexcept {
+    IPC_CONSTEXPR_ static std::size_t block_size(std::size_t id) noexcept {
-        return base_size + (value + 1) * iter_size;
+        return (id < classes_size) ? (base_size + (id + 1) * iter_size) : 0;
+    }
+
+    template <typename F, typename D, typename ... P>
+    IPC_CONSTEXPR_ static auto classify(F f, D d, std::size_t size, P ... params) {
+        std::size_t id = (size - base_size - 1) / iter_size;
+        return (id < classes_size) ? f(id, params..., size) : d(params..., size);
     }
 };
 
-template <std::size_t B, std::size_t I>
+template <typename FixedAlloc,
-const std::size_t default_mapping_policy<B, I>::table[default_mapping_policy<B, I>::classes_size] = {
+          typename DefaultAlloc = mem::static_alloc,
-    /* 1 - 8 ~ 32 */
+          typename MappingP     = default_mapping_policy<>>
-    0 , 1 , 2 , 3 ,
+class static_variable_wrapper {
-    /* 2 - 48 ~ 256 */
+private:
-    5 , 5 , 7 , 7 , 9 , 9 , 11, 11, 13, 13, 15, 15, 17, 17,
+    static FixedAlloc& instance(std::size_t id) {
-    19, 19, 21, 21, 23, 23, 25, 25, 27, 27, 29, 29, 31, 31
+        static struct initiator {
-};
+            std::aligned_storage_t<sizeof (FixedAlloc), 
-
+                                   alignof(FixedAlloc)> arr_[MappingP::classes_size];
-template <template <std::size_t> class Fixed,
+            initiator() {
-          typename MappingP    = default_mapping_policy<>,
+                MappingP::foreach([](std::size_t id, initiator* t) {
-          typename StaticAlloc = mem::static_alloc>
+                    ::new (&(t->arr_[id])) FixedAlloc(MappingP::block_size(id));
-class variable_wrapper {
+                }, this);
-
+            }
-    template <typename F>
+        } init__;
-    constexpr static auto choose(std::size_t size, F&& f) {
+        return reinterpret_cast<FixedAlloc&>(init__.arr_[id]);
-        return ipc::detail::static_switch<MappingP::classes_size>(MappingP::classify(size), [&f](auto index) {
-            return f(Fixed<MappingP::block_size(decltype(index)::value)>{});
-        }, [&f] {
-            return f(StaticAlloc{});
-        });
     }
 
 public:
-    static void swap(variable_wrapper&) {}
+    static void swap(static_variable_wrapper&) {}
-
-    static void clear() {
-        ipc::detail::static_for<MappingP::classes_size>([](auto index) {
-            Fixed<MappingP::block_size(decltype(index)::value)>::clear();
-        });
-        StaticAlloc::clear();
-    }
 
     static void* alloc(std::size_t size) {
-        return choose(size, [size](auto&& alc) { return alc.alloc(size); });
+        return MappingP::classify([](std::size_t id, std::size_t size) {
+            return instance(id).alloc(size);
+        }, DefaultAlloc::alloc, size);
     }
 
     static void free(void* p, std::size_t size) {
-        choose(size, [p, size](auto&& alc) { alc.free(p, size); });
+        MappingP::classify([](std::size_t id, void* p, std::size_t size) {
+            instance(id).free(p, size);
+        }, static_cast<void(*)(void*, std::size_t)>(DefaultAlloc::free), size, p);
     }
 };
 

src/pool_alloc.cpp

@@ -5,10 +5,6 @@
 namespace ipc {
 namespace mem {
 
-void pool_alloc::clear() {
-    async_pool_alloc::clear();
-}
-
 void* pool_alloc::alloc(std::size_t size) {
     return async_pool_alloc::alloc(size);
 }

test/test.h

@@ -40,12 +40,11 @@ struct test_stopwatch {
         }
     }
 
-    template <int Factor>
+    template <int Factor, typename ToDur = std::chrono::microseconds>
-    void print_elapsed(int N, int M, int Loops) {
+    void print_elapsed(int N, int M, int Loops, const char * unit = " us/d") {
-        auto ts = sw_.elapsed<std::chrono::microseconds>();
+        auto ts = sw_.elapsed<ToDur>();
         std::cout << "[" << N << ":" << M << ", " << Loops << "] "
-                  << "performance: " << (ts / 1000.0) << " ms, "
+                  << (double(ts) / double(Factor ? (Loops * Factor) : (Loops * N))) << unit << std::endl;
-                  << (double(ts) / double(Factor ? (Loops * Factor) : (Loops * N))) << " us/d" << std::endl;
     }
 
     void print_elapsed(int N, int M, int Loops) {

test/test_circ.cpp

@@ -238,7 +238,8 @@ private slots:
     void test_prod_cons_1v3();
     void test_prod_cons_performance();
     void test_queue();
-} unit__;
+};
+// } unit__;
 
 #include "test_circ.moc"
 

test/test_ipc.cpp

@@ -30,7 +30,7 @@ std::vector<ipc::buff_t> datas__;
 constexpr int DataMin   = 2;
 constexpr int DataMax   = 256;
 constexpr int LoopCount = 100000;
-//constexpr int LoopCount = 1000;
+// constexpr int LoopCount = 1000;
 
 } // internal-linkage
 
@@ -179,6 +179,7 @@ private slots:
     void test_channel();
     void test_channel_rtt();
     void test_channel_performance();
+// };
 } unit__;
 
 #include "test_ipc.moc"
@@ -352,7 +353,6 @@ void Unit::test_route() {
 }
 
 void Unit::test_route_rtt() {
-    //return;
     test_stopwatch sw;
 
     std::thread t1 {[&] {
@@ -392,15 +392,13 @@ void Unit::test_route_rtt() {
 }
 
 void Unit::test_route_performance() {
-    //return;
     ipc::detail::static_for<8>([](auto index) {
         test_prod_cons<ipc::route, 1, decltype(index)::value + 1, false>();
     });
-    test_prod_cons<ipc::route, 1, 8>(); // test & verify
+    // test_prod_cons<ipc::route, 1, 8>(); // test & verify
 }
 
 void Unit::test_channel() {
-    //return;
     std::thread t1 {[&] {
         ipc::channel cc { "my-ipc-channel" };
         for (std::size_t i = 0;; ++i) {

test/test_mem.cpp

@@ -1,4 +1,5 @@
 #include <vector>
+#include <array>
 #include <thread>
 #include <atomic>
 #include <cstddef>
@@ -8,7 +9,7 @@
 #include "memory/resource.h"
 #include "pool_alloc.h"
 
-//#include "gperftools/tcmalloc.h"
+// #include "gperftools/tcmalloc.h"
 
 #include "test.h"
 
@@ -24,58 +25,77 @@ class Unit : public TestSuite {
 private slots:
     void initTestCase();
 
-    void test_alloc_free();
+    void test_static_alloc();
-} unit__;
+    void test_pool_alloc();
+    void test_tc_alloc();
+};
+// } unit__;
 
 #include "test_mem.moc"
 
 constexpr int DataMin   = 4;
 constexpr int DataMax   = 256;
-constexpr int LoopCount = 100000;
+constexpr int LoopCount = 4194304;
+
+// constexpr int DataMin   = 256;
+// constexpr int DataMax   = 512;
+// constexpr int LoopCount = 2097152;
 
 std::vector<std::size_t> sizes__;
 
 template <typename M>
 struct alloc_ix_t {
-    static std::vector<int> ix_[2];
+    static std::vector<int> ix_;
     static bool inited_;
+
+    alloc_ix_t() {
+        if (inited_) return;
+        inited_ = true;
+        M::init(ix_);
+    }
+
+    int index(std::size_t /*pid*/, std::size_t /*k*/, std::size_t n) {
+        return ix_[n];
+    }
 };
 
 template <typename M>
-std::vector<int> alloc_ix_t<M>::ix_[2] = { std::vector<int>(LoopCount), std::vector<int>(LoopCount) };
+std::vector<int> alloc_ix_t<M>::ix_(LoopCount);
 template <typename M>
 bool alloc_ix_t<M>::inited_ = false;
 
-struct alloc_random : alloc_ix_t<alloc_random> {
+template <std::size_t N>
-    alloc_random() {
+struct alloc_FIFO : alloc_ix_t<alloc_FIFO<N>> {
-        if (inited_) return;
+    static void init(std::vector<int>& ix) {
-        inited_ = true;
+        for (int i = 0; i < LoopCount; ++i) {
+            ix[static_cast<std::size_t>(i)] = i;
+        }
+    }
+};
+
+template <std::size_t N>
+struct alloc_LIFO : alloc_ix_t<alloc_LIFO<N>> {
+    static void init(std::vector<int>& ix) {
+        for (int i = 0; i < LoopCount; ++i) {
+            ix[static_cast<std::size_t>(i)] = i;
+        }
+    }
+
+    int index(std::size_t pid, std::size_t k, std::size_t n) {
+        constexpr static int CacheSize = LoopCount / N;
+        if (k) {
+            return this->ix_[(CacheSize * (2 * pid + 1)) - 1 - n];
+        }
+        else return this->ix_[n];
+    }
+};
+
+template <std::size_t N>
+struct alloc_random : alloc_ix_t<alloc_random<N>> {
+    static void init(std::vector<int>& ix) {
         capo::random<> rdm_index(0, LoopCount - 1);
         for (int i = 0; i < LoopCount; ++i) {
-            ix_[0][static_cast<std::size_t>(i)] =
+            ix[static_cast<std::size_t>(i)] = rdm_index();
-            ix_[1][static_cast<std::size_t>(i)] = rdm_index();
-        }
-    }
-};
-
-struct alloc_LIFO : alloc_ix_t<alloc_LIFO> {
-    alloc_LIFO() {
-        if (inited_) return;
-        inited_ = true;
-        for (int i = 0, n = LoopCount - 1; i < LoopCount; ++i, --n) {
-            ix_[0][static_cast<std::size_t>(i)] =
-            ix_[1][static_cast<std::size_t>(n)] = i;
-        }
-    }
-};
-
-struct alloc_FIFO : alloc_ix_t<alloc_FIFO> {
-    alloc_FIFO() {
-        if (inited_) return;
-        inited_ = true;
-        for (int i = 0; i < LoopCount; ++i) {
-            ix_[0][static_cast<std::size_t>(i)] =
-            ix_[1][static_cast<std::size_t>(i)] = i;
         }
     }
 };
@@ -89,33 +109,13 @@ void Unit::initTestCase() {
     }
 }
 
-template <typename AllocT>
+template <typename AllocT, int ThreadsN>
-void benchmark_alloc() {
-    std::cout << std::endl << type_name<AllocT>() << std::endl;
-
-    test_stopwatch sw;
-    sw.start();
-
-    for (std::size_t k = 0; k < 100; ++k)
-    for (std::size_t n = 0; n < LoopCount; ++n) {
-        std::size_t s = sizes__[n];
-        AllocT::free(AllocT::alloc(s), s);
-    }
-
-    sw.print_elapsed<1>(DataMin, DataMax, LoopCount * 100);
-}
-
-template <typename AllocT, typename ModeT, int ThreadsN>
 void benchmark_alloc() {
     std::cout << std::endl
-              << "[Threads: " << ThreadsN << ", Mode: " << type_name<ModeT>() << "] "
+              << "[Threads: " << ThreadsN << "] "
               << type_name<AllocT>() << std::endl;
 
-    std::vector<void*> ptrs[ThreadsN];
+    constexpr static int CacheSize = LoopCount / ThreadsN;
-    for (auto& vec : ptrs) {
-        vec.resize(LoopCount);
-    }
-    ModeT mode;
 
     std::atomic_int fini { 0 };
     test_stopwatch sw;
@@ -126,11 +126,50 @@ void benchmark_alloc() {
     for (auto& w : works) {
         w = std::thread {[&, pid] {
             sw.start();
-            for (std::size_t k = 0; k < 10; ++k)
+            for (std::size_t k = 0; k < 100; ++k)
-            for (std::size_t x = 0; x < 2; ++x) {
+            for (std::size_t n = (CacheSize * pid); n < (CacheSize * (pid + 1)); ++n) {
-                for(std::size_t n = 0; n < LoopCount; ++n) {
+                std::size_t s = sizes__[n];
-                    int    m = mode.ix_[x][n];
+                AllocT::free(AllocT::alloc(s), s);
-                    void*& p = ptrs[pid][static_cast<std::size_t>(m)];
+            }
+            if ((fini.fetch_add(1, std::memory_order_relaxed) + 1) == ThreadsN) {
+                sw.print_elapsed<1, std::chrono::nanoseconds>(DataMin, DataMax, LoopCount * 100, " ns/d");
+            }
+        }};
+        ++pid;
+    }
+
+    for (auto& w : works) w.join();
+}
+
+template <typename AllocT, template <std::size_t> class ModeT, int ThreadsN>
+void benchmark_alloc() {
+    std::cout << std::endl
+              << "[Threads: " << ThreadsN << ", Mode: " << type_name<ModeT<ThreadsN>>() << "] "
+              << type_name<AllocT>() << std::endl;
+
+    constexpr static int CacheSize = LoopCount / ThreadsN;
+
+    std::vector<void*> ptrs[ThreadsN];
+    for (auto& vec : ptrs) {
+        vec.resize(LoopCount);
+    }
+
+    ModeT<ThreadsN> mode;
+
+    std::atomic_int fini { 0 };
+    test_stopwatch  sw;
+
+    std::thread works[ThreadsN];
+    int pid = 0;
+
+    for (auto& w : works) {
+        w = std::thread {[&, pid] {
+            auto& vec = ptrs[pid];
+            sw.start();
+            for (std::size_t k = 0; k < 2; ++k)
+            for (std::size_t n = (CacheSize * pid); n < (CacheSize * (pid + 1)); ++n) {
+                int    m = mode.index(pid, k, n);
+                void*& p = vec[static_cast<std::size_t>(m)];
                 std::size_t s  = sizes__[static_cast<std::size_t>(m)];
                 if (p == nullptr) {
                     p = AllocT::alloc(s);
@@ -140,35 +179,50 @@ void benchmark_alloc() {
                     p = nullptr;
                 }
             }
-            }
             if ((fini.fetch_add(1, std::memory_order_relaxed) + 1) == ThreadsN) {
-                sw.print_elapsed<1>(DataMin, DataMax, LoopCount * 10 * ThreadsN);
+                sw.print_elapsed<1>(DataMin, DataMax, LoopCount);
             }
         }};
         ++pid;
     }
-    sw.start();
 
     for (auto& w : works) w.join();
 }
 
-template <typename AllocT, typename ModeT, int ThreadsN>
+template <typename AllocT, template <std::size_t> class ModeT, int ThreadsN>
 struct test_performance {
     static void start() {
-        test_performance<AllocT, ModeT, ThreadsN - 1>::start();
+        test_performance<AllocT, ModeT, ThreadsN / 2>::start();
         benchmark_alloc<AllocT, ModeT, ThreadsN>();
     }
 };
 
-template <typename AllocT, typename ModeT>
+template <typename AllocT, template <std::size_t> class ModeT>
 struct test_performance<AllocT, ModeT, 1> {
     static void start() {
         benchmark_alloc<AllocT, ModeT, 1>();
     }
 };
 
-//class tc_alloc {
+template <std::size_t> struct dummy;
-//public:
+
+template <typename AllocT, int ThreadsN>
+struct test_performance<AllocT, dummy, ThreadsN> {
+    static void start() {
+        test_performance<AllocT, dummy, ThreadsN / 2>::start();
+        benchmark_alloc<AllocT, ThreadsN>();
+    }
+};
+
+template <typename AllocT>
+struct test_performance<AllocT, dummy, 1> {
+    static void start() {
+        benchmark_alloc<AllocT, 1>();
+    }
+};
+
+// class tc_alloc {
+// public:
 //    static void clear() {}
 
 //    static void* alloc(std::size_t size) {
@@ -178,18 +232,30 @@ struct test_performance<AllocT, ModeT, 1> {
 //    static void free(void* p, std::size_t size) {
 //        tc_free_sized(p, size);
 //    }
-//};
+// };
 
-#define TEST_ALLOC_TYPE /*ipc::mem::static_alloc*/ ipc::mem::async_pool_alloc /*tc_alloc*/
+void Unit::test_static_alloc() {
+    // test_performance<ipc::mem::static_alloc, dummy       , 128>::start();
+    // test_performance<ipc::mem::static_alloc, alloc_FIFO  , 128>::start();
+    // test_performance<ipc::mem::static_alloc, alloc_LIFO  , 128>::start();
+    // test_performance<ipc::mem::static_alloc, alloc_random, 128>::start();
+}
 
-void Unit::test_alloc_free() {
+void Unit::test_pool_alloc() {
-//    benchmark_alloc <TEST_ALLOC_TYPE>();
+    test_performance<ipc::mem::async_pool_alloc, dummy       , 128>::start();
-//    test_performance<TEST_ALLOC_TYPE, alloc_FIFO  , 24>::start();
+    test_performance<ipc::mem::async_pool_alloc, alloc_FIFO  , 128>::start();
 
-    benchmark_alloc <TEST_ALLOC_TYPE>();
+    test_performance<ipc::mem::async_pool_alloc, dummy       , 128>::start();
-    test_performance<TEST_ALLOC_TYPE, alloc_FIFO  , 16>::start();
+    test_performance<ipc::mem::async_pool_alloc, alloc_FIFO  , 128>::start();
-    test_performance<TEST_ALLOC_TYPE, alloc_LIFO  , 16>::start();
+    test_performance<ipc::mem::async_pool_alloc, alloc_LIFO  , 128>::start();
-    test_performance<TEST_ALLOC_TYPE, alloc_random, 16>::start();
+    test_performance<ipc::mem::async_pool_alloc, alloc_random, 128>::start();
+}
+
+void Unit::test_tc_alloc() {
+    // test_performance<tc_alloc, dummy       , 128>::start();
+    // test_performance<tc_alloc, alloc_FIFO  , 128>::start();
+    // test_performance<tc_alloc, alloc_LIFO  , 128>::start();
+    // test_performance<tc_alloc, alloc_random, 128>::start();
 }
 
 } // internal-linkage