diff --git a/dipu/torch_dipu/csrc_dipu/base/DIPUGlobals.cpp b/dipu/torch_dipu/csrc_dipu/base/DIPUGlobals.cpp
index c214a17c3..22fa32b01 100644
--- a/dipu/torch_dipu/csrc_dipu/base/DIPUGlobals.cpp
+++ b/dipu/torch_dipu/csrc_dipu/base/DIPUGlobals.cpp
@@ -52,7 +52,7 @@ void releaseAllResourcesImpl() {
   called = true;
   releaseAllGenerator();
   releaseAllDeviceMem();
-  releaseAllEvent();
+  event_pool_clear();
   devproxy::finalizeVendor();
 }
 
diff --git a/dipu/torch_dipu/csrc_dipu/base/basedef.h b/dipu/torch_dipu/csrc_dipu/base/basedef.h
index 8f416e8ef..8e76117f7 100644
--- a/dipu/torch_dipu/csrc_dipu/base/basedef.h
+++ b/dipu/torch_dipu/csrc_dipu/base/basedef.h
@@ -35,3 +35,7 @@ const auto DIPU_BACKEND_SPARSE_TYPE =
 const auto DICL_BACKEND_NAME = "dicl";
 
 }  // namespace dipu
+
+namespace dipu {
+auto inline constexpr kMaxDeviceNumber = 16;
+}
diff --git a/dipu/torch_dipu/csrc_dipu/base/utility.h b/dipu/torch_dipu/csrc_dipu/base/utility.h
new file mode 100644
index 000000000..0b0db74ca
--- /dev/null
+++ b/dipu/torch_dipu/csrc_dipu/base/utility.h
@@ -0,0 +1,53 @@
+#pragma once
+
+#include <array>
+#include <utility>
+
+namespace dipu {
+
+template <typename T, typename S>
+struct make_static_value_array {};
+
+template <typename T, std::size_t... I>
+struct make_static_value_array<T, std::index_sequence<I...>> {
+  using value_type = std::decay_t<decltype(T::template value<0>)>;
+  using type = std::array<value_type, sizeof...(I)>;
+  type static inline constexpr value{T::template value<I>...};
+};
+
+// Create an array of static values. Each type T should provide a static
+// 'value()' function to return a reference of the underlying static variable.
+//
+// In this way, value can be lazily initialized.
+template <typename T, std::size_t N>
+// Note: in C++17 "inline constexpr" (without "static") is necessary to make
+// sure there is only one entity.
+auto inline constexpr make_static_value_array_v =
+    make_static_value_array<T, std::make_index_sequence<N>>::value;
+
+// A CRTP helper to simplify access method of static value holder.
+//
+// Usage:
+//
+// ```
+// struct integer : static_value_array<foo, 16> {
+//   template <std::size_t I> auto static value() -> int & {
+//     auto static instance = 0;
+//     return instance;
+//   }
+// };
+//
+// Then we can use `integer::get(1)` to access `integer::value<1>()`.
+template <typename T, std::size_t N>
+struct static_value_array {
+  auto static array() -> decltype(auto) {
+    auto static constexpr instance = make_static_value_array_v<T, N>;
+    return instance;
+  }
+  template <typename... A>
+  auto static get(std::size_t index, A&&... args) -> decltype(auto) {
+    return array()[index](std::forward<A>(args)...);
+  }
+};
+
+}  // namespace dipu
diff --git a/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.cpp b/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.cpp
index d3ad103fa..5ff3d11b9 100644
--- a/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.cpp
+++ b/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.cpp
@@ -1,102 +1,77 @@
 #include "DIPUEventPool.h"
 
 #include <deque>
-#include <functional>
-#include <iostream>
 #include <mutex>
 
-namespace dipu {
+#include "csrc_dipu/base/basedef.h"
+#include "csrc_dipu/base/utility.h"
+#include "csrc_dipu/runtime/device/deviceapis.h"
 
-template <typename T>
-class EventPool final {
- protected:
-  std::deque<T> event_pool_;
-  unsigned int allocate_num_ = 0;
+namespace {
 
-  std::function<void(T&)> allocator_;
-  std::function<void(T&)> deleter_;
-  using mutex_t = std::recursive_mutex;
-  mutex_t event_mutex_;
+class EventPool {
+  std::deque<dipu::deviceEvent_t> pool;
+  std::recursive_mutex mutex;
 
  public:
-  EventPool(const std::function<void(T&)>& allocator,
-            const std::function<void(T&)>& deleter)
-      : allocator_(allocator), deleter_(deleter) {}
-
-  EventPool(const EventPool&) = delete;
-  EventPool(EventPool&&) = delete;
-  EventPool& operator=(const EventPool&) = delete;
-  EventPool& operator=(EventPool&&) = delete;
-
-  ~EventPool() = default;
-
-  void release() {
-    std::lock_guard<mutex_t> _(event_mutex_);
-    for (auto& event : event_pool_) {
-      deleter_(event);
-      allocate_num_--;
-    }
-    event_pool_.clear();
-  }
-
-  void get(T& event) {
-    bool need_allocator = false;
+  void acquire(dipu::deviceEvent_t& event) {
     {
-      std::lock_guard<mutex_t> _(event_mutex_);
-      if (event_pool_.empty()) {
-        need_allocator = true;
-      } else {
-        event = event_pool_.back();
-        event_pool_.pop_back();
+      std::scoped_lock _(mutex);
+      if (!pool.empty()) {
+        event = pool.back();
+        pool.pop_back();
+        return;
       }
     }
-    if (need_allocator) {
-      allocator_(event);
-    }
+
+    dipu::devapis::createEvent(&event);
+  }
+
+  void release(dipu::deviceEvent_t& event) {
+    std::scoped_lock _(mutex);
+    pool.emplace_back(event);
   }
 
-  void restore(T& event) {
-    std::lock_guard<mutex_t> _(event_mutex_);
-    event_pool_.emplace_back(event);
+  void clear() {  // should it called inside destructor?
+    std::scoped_lock _(mutex);
+    for (auto& event : pool) {
+      dipu::devapis::destroyEvent(event);
+    }
+    pool.clear();
   }
 };
 
-EventPool<deviceEvent_t>* getEventPool() {
-  const int index = devproxy::current_device();
-// GlobalEventPool for different cards , construct when really needed
-#define dispatch_event_pool(device_id)                               \
-  if (index == (device_id)) {                                        \
-    static EventPool<deviceEvent_t> gDIPUEventPool(                  \
-        [](deviceEvent_t& event) { devapis::createEvent(&event); },  \
-        [](deviceEvent_t& event) { devapis::destroyEvent(event); }); \
-    return &gDIPUEventPool;                                          \
+struct StaticEventPoolArray
+    : dipu::static_value_array<StaticEventPoolArray, dipu::kMaxDeviceNumber> {
+  template <std::size_t I>
+  inline static auto& value() {
+    auto static instance = EventPool();
+    return instance;
   }
+};
 
-  dispatch_event_pool(0);
-  dispatch_event_pool(1);
-  dispatch_event_pool(2);
-  dispatch_event_pool(3);
-  dispatch_event_pool(4);
-  dispatch_event_pool(5);
-  dispatch_event_pool(6);
-  dispatch_event_pool(7);
-  dispatch_event_pool(8);
-  dispatch_event_pool(9);
-  dispatch_event_pool(10);
-  dispatch_event_pool(11);
-  dispatch_event_pool(12);
-  dispatch_event_pool(13);
-  dispatch_event_pool(14);
-  dispatch_event_pool(15);
-  TORCH_CHECK(false, "support up to 16 cards");
+auto event_pool(int index) -> EventPool& {
+  TORCH_CHECK(0 <= index and index < dipu::kMaxDeviceNumber,
+              "device index out of range");
+  return StaticEventPoolArray::get(index);
 }
 
-void getEventFromPool(deviceEvent_t& event) { getEventPool()->get(event); }
+}  // namespace
+
+namespace dipu {
+
+void event_pool_acquire(int index, deviceEvent_t& event) {
+  event_pool(index).acquire(event);
+}
 
-void restoreEventToPool(deviceEvent_t& event) {
-  getEventPool()->restore(event);
+void event_pool_release(int index, deviceEvent_t& event) {
+  event_pool(index).release(event);
 }
 
-void releaseAllEvent() { getEventPool()->release(); }
+void event_pool_clear() {
+  for (auto& getter : StaticEventPoolArray::array()) {
+    getter().clear();
+  }
+}
 
 }  // namespace dipu
diff --git a/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.h b/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.h
index 1dd105ac1..e8aceaac2 100644
--- a/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.h
+++ b/dipu/torch_dipu/csrc_dipu/runtime/core/DIPUEventPool.h
@@ -6,10 +6,8 @@
 
 namespace dipu {
 
-void getEventFromPool(deviceEvent_t& event);
-
-void restoreEventToPool(deviceEvent_t& event);
-
-void releaseAllEvent();
+void event_pool_acquire(int index, deviceEvent_t& event);
+void event_pool_release(int index, deviceEvent_t& event);
+void event_pool_clear();
 
 }  // namespace dipu
diff --git a/dipu/torch_dipu/csrc_dipu/runtime/core/allocator/DIPUCachingAllocator.h b/dipu/torch_dipu/csrc_dipu/runtime/core/allocator/DIPUCachingAllocator.h
index 6d557ba41..e1b9d309b 100644
--- a/dipu/torch_dipu/csrc_dipu/runtime/core/allocator/DIPUCachingAllocator.h
+++ b/dipu/torch_dipu/csrc_dipu/runtime/core/allocator/DIPUCachingAllocator.h
@@ -5,6 +5,8 @@
 #include <c10/core/Device.h>
 #include <c10/util/flat_hash_map.h>
 
+#include "csrc_dipu/base/basedef.h"
+#include "csrc_dipu/base/utility.h"
 #include "csrc_dipu/runtime/core/DIPUEvent.h"
 
 #include "DIPUAsyncResourcePool.h"
@@ -213,48 +215,33 @@ struct RawAllocator<at::DeviceType::CPU> {
   using type = DIPURawHostAllocator;
 };
 
-template <typename AllocatorImpl, class AsyncMemPoolImpl, int device_id>
-c10::Allocator* get_allocator_impl(c10::Allocator* raw_allocator) {
-  // Construct when really needed
-  // async_mem_pool is used when cache_allocator being destructed so it should
-  // be destructed after cache_allocator
-  static AsyncMemPoolImpl async_mem_pool;
-  static AllocatorImpl cache_allocator;
-  static int n = [&]() {
-    cache_allocator.set_raw_allocator(raw_allocator);
-    cache_allocator.set_async_mem_pool(&async_mem_pool);
-    return 0;
-  }();
-  return &cache_allocator;
-}
-
-template <class AllocatorImpl, class AsyncMemPoolImpl>
-c10::Allocator* get_allocator(int device_id, c10::Allocator* raw_allocator) {
-#define DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(id)                       \
-  if (device_id == (id)) {                                          \
-    return get_allocator_impl<AllocatorImpl, AsyncMemPoolImpl, id>( \
-        raw_allocator);                                             \
+template <typename AllocatorImpl, class AsyncMemPoolImpl>
+struct StaticAllocatorArray
+    : dipu::static_value_array<
+          StaticAllocatorArray<AllocatorImpl, AsyncMemPoolImpl>,
+          dipu::kMaxDeviceNumber> {
+  template <std::size_t I>
+  inline static c10::Allocator* value(c10::Allocator* raw_allocator) {
+    // Construct when really needed
+    // async_mem_pool is used when cache_allocator being destructed so it should
+    // be destructed after cache_allocator
+    static AsyncMemPoolImpl async_mem_pool;
+    static AllocatorImpl cache_allocator;
+    static int n = [&]() {
+      cache_allocator.set_raw_allocator(raw_allocator);
+      cache_allocator.set_async_mem_pool(&async_mem_pool);
+      return 0;
+    }();
+    return &cache_allocator;
   }
+};
 
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(0);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(1);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(2);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(3);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(4);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(5);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(6);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(7);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(8);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(9);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(10);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(11);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(12);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(13);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(14);
-  DIPU_ALLOCATOR_DISPATCH_DEVICE_ID(15);
-  TORCH_CHECK(false, "support up to 16 cards");
+template <class Allocator, class MemPool>
+c10::Allocator* get_allocator(int index, c10::Allocator* raw_allocator) {
+  TORCH_CHECK(0 <= index and index < dipu::kMaxDeviceNumber,
+              "device index out of range");
+  return StaticAllocatorArray<Allocator, MemPool>::get(index, raw_allocator);
 }
-#undef DIPU_ALLOCATOR_DISPATCH_DEVICE_ID
 
 #define DIPU_REGISTER_ALLOCATOR(name, device_type, CachingAllocator,          \
                                 algorithm, priority)                          \
diff --git a/dipu/torch_dipu/csrc_dipu/runtime/devproxy/deviceproxy.cpp b/dipu/torch_dipu/csrc_dipu/runtime/devproxy/deviceproxy.cpp
index e2a460156..e8ff561d8 100644
--- a/dipu/torch_dipu/csrc_dipu/runtime/devproxy/deviceproxy.cpp
+++ b/dipu/torch_dipu/csrc_dipu/runtime/devproxy/deviceproxy.cpp
@@ -172,9 +172,15 @@ bool isStreamEmpty(deviceStream_t stream) {
 //  device event related
 // =====================
 
-void createEvent(deviceEvent_t* event) { return getEventFromPool(*event); }
+void createEvent(deviceEvent_t* event) {
+  auto index = current_device();
+  return event_pool_acquire(index, *event);
+}
 
-void destroyEvent(deviceEvent_t event) { return restoreEventToPool(event); }
+void destroyEvent(deviceEvent_t event) {
+  auto index = current_device();
+  return event_pool_release(index, event);
+}
 
 void waitEvent(deviceEvent_t event) { return devapis::waitEvent(event); }