diff --git a/.gitmodules b/.gitmodules
index 0de60d4c80e7..498ba35b6d15 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,9 +1,9 @@
 [submodule "mshadow"]
 	path = mshadow
-	url = /~https://github.com/dmlc/mshadow.git
+	url = /~https://github.com/dato-code/mshadow.git
 [submodule "dmlc-core"]
 	path = dmlc-core
-	url = /~https://github.com/dmlc/dmlc-core.git
+	url = /~https://github.com/dato-code/dmlc-core.git
 [submodule "ps-lite"]
 	path = ps-lite
 	url = /~https://github.com/dmlc/ps-lite
diff --git a/Makefile b/Makefile
index 4e95f85ea321..7de7c388688f 100644
--- a/Makefile
+++ b/Makefile
@@ -12,15 +12,10 @@ ifndef DMLC_CORE
 	DMLC_CORE = dmlc-core
 endif
 
-ifneq ($(USE_OPENMP), 1)
-	export NO_OPENMP = 1
-endif
-
 # use customized config file
 include $(config)
 include mshadow/make/mshadow.mk
 include $(DMLC_CORE)/make/dmlc.mk
-unexport NO_OPENMP
 
 # all the possible warning tread
 WARNFLAGS= -Wall
@@ -32,7 +27,16 @@ ifeq ($(DEBUG), 1)
 else
 	CFLAGS += -O3
 endif
-CFLAGS += -I./mshadow/ -I./dmlc-core/include -fPIC -Iinclude $(MSHADOW_CFLAGS)
+
+ifneq ($(WIN32), 1)
+	CFLAGS += -fPIC
+	SHARED_LIB_EXT = so
+	STATIC_LIB_EXT = a
+else
+	SHARED_LIB_EXT = dll
+	STATIC_LIB_EXT = lib
+endif
+CFLAGS += -I./mshadow/ -I./dmlc-core/include -Iinclude $(MSHADOW_CFLAGS)
 LDFLAGS = -pthread $(MSHADOW_LDFLAGS) $(DMLC_LDFLAGS)
 ifeq ($(DEBUG), 1)
 	NVCCFLAGS = -g -G -O0 -ccbin $(CXX) $(MSHADOW_NVCCFLAGS)
@@ -96,7 +100,7 @@ include $(MXNET_PLUGINS)
 
 .PHONY: clean all test lint doc clean_all rcpplint rcppexport roxygen
 
-all: lib/libmxnet.a lib/libmxnet.so $(BIN)
+all: lib/libmxnet.$(STATIC_LIB_EXT) lib/libmxnet.$(SHARED_LIB_EXT) $(BIN)
 
 SRC = $(wildcard src/*.cc src/*/*.cc)
 OBJ = $(patsubst %.cc, build/%.o, $(SRC))
@@ -177,19 +181,19 @@ $(EXTRA_OPERATORS)/build/%_gpu.o: $(EXTRA_OPERATORS)/%.cu
 	$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CFLAGS) -Isrc/operator" -M -MT $(EXTRA_OPERATORS)/build/$*_gpu.o $< >$(EXTRA_OPERATORS)/build/$*_gpu.d
 	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS) -Isrc/operator" $<
 
-# NOTE: to statically link libmxnet.a we need the option
-# --Wl,--whole-archive -lmxnet --Wl,--no-whole-archive
-lib/libmxnet.a: $(ALL_DEP)
-	@mkdir -p $(@D)
-	ar crv $@ $(filter %.o, $?)
-
 copy_cuda_deps:
 ifdef CUDA_DEP
 	@mkdir -p lib
 	cp $(CUDA_DEP) lib
 endif
 
-lib/libmxnet.so: $(ALL_DEP) copy_cuda_deps
+# NOTE: to statically link libmxnet.a we need the option
+# --Wl,--whole-archive -lmxnet --Wl,--no-whole-archive
+lib/libmxnet.$(STATIC_LIB_EXT): $(ALL_DEP)
+	@mkdir -p $(@D)
+	ar crv $@ $(filter %.o, $?)
+
+lib/libmxnet.$(SHARED_LIB_EXT): $(ALL_DEP) copy_cuda_deps
 	@mkdir -p $(@D)
 	$(CXX) $(CFLAGS) -shared -o $@ $(filter %.o %.a, $^) $(LDFLAGS)
 
diff --git a/configure b/configure
index 292476a58284..bd82450128a1 100755
--- a/configure
+++ b/configure
@@ -242,15 +242,19 @@ CXX=$CXX
 # OS Dependent Flags
 if [[ $OSTYPE == linux* ]]; then
         USE_OPENMP=1
+        WIN32=0
         USE_BLAS="blas"
         SHARED_LINKER_FLAGS="-Wl,-rpath,\\\$\$ORIGIN -Wl,-rpath,${PWD}/deps/local/lib64 -Wl,-rpath,${PWD}/deps/local/lib"
 elif [[ $OSTYPE == darwin* ]]; then
         USE_OPENMP=0
+        WIN32=0
         USE_BLAS="apple"
         SHARED_LINKER_FLAGS="-Wl,-rpath,@loader_path -Wl,-rpath,${PWD}/deps/local/lib64 -Wl,-rpath,${PWD}/deps/local/lib"
 elif [[ $OSTYPE == msys ]]; then
         USE_OPENMP=0
-        USE_BLAS="blas"
+        WIN32=1
+        USE_BLAS="openblas"
+        ADD_CFLAGS="-DMXNET_EXPORTS "
         SHARED_LINKER_FLAGS="-Wl,-rpath,/mingw64/bin -Wl,-rpath,${PWD}/deps/local/lib64 -Wl,-rpath,${PWD}/deps/local/lib"
 fi
 
diff --git a/dmlc-core b/dmlc-core
index ea9b247b6f99..2d9119a06b1f 160000
--- a/dmlc-core
+++ b/dmlc-core
@@ -1 +1 @@
-Subproject commit ea9b247b6f9965c95aa66f42374d0867c46d9abd
+Subproject commit 2d9119a06b1fed465cdc6679cf76557d5964e23d
diff --git a/include/mxnet/base.h b/include/mxnet/base.h
index 52100cdf05ea..f9880a20b2cc 100644
--- a/include/mxnet/base.h
+++ b/include/mxnet/base.h
@@ -52,7 +52,7 @@
 /*!
  * \brief define dllexport for Visual Studio
  */
-#ifdef _MSC_VER
+#if defined(_MSC_VER) || defined(_WIN32) 
 #ifdef MXNET_EXPORTS
 #define MXNET_API __declspec(dllexport)
 #else
diff --git a/include/mxnet/engine.h b/include/mxnet/engine.h
index 011beef5c7b9..529d9d01b2bf 100644
--- a/include/mxnet/engine.h
+++ b/include/mxnet/engine.h
@@ -171,7 +171,9 @@ class MXNET_API Engine {
   /*!\brief virtual destructor */
   virtual ~Engine() noexcept(false) {}
   /*!
-   * \return Engine singleton.
+   * \return Engine singleton. NULL if Shutdown() is called.
+   *
+   * \note Check nullptr if used in destructors.
    */
   static Engine* Get();
   /*!
@@ -180,9 +182,15 @@ class MXNET_API Engine {
    *  This function is called by another singleton X who requires
    *  engine to be destructed after X.
    *
-   * \return A shared pointer to Engine singleton.
+   * \return A ref of the shared pointer to Engine singleton.
+   */
+  static std::shared_ptr<Engine>& _GetSharedRef();
+
+  /**
+   * \brief Destroy the global engine singleton. Future
+   * access to engine is disabled.
    */
-  static std::shared_ptr<Engine> _GetSharedRef();
+  static void Shutdown();
   /*!
    * \brief Push an synchronous operation to the engine.
    * \param exec_fn Execution function that executes the operation.
diff --git a/include/mxnet/ndarray.h b/include/mxnet/ndarray.h
index 811fa2c968a4..60ffd18b5db2 100644
--- a/include/mxnet/ndarray.h
+++ b/include/mxnet/ndarray.h
@@ -323,13 +323,18 @@ class NDArray {
     }
     /*! \brief destructor */
     ~Chunk() {
+      auto engine_ptr = Engine::Get();
       if (static_data || delay_alloc) {
-        Engine::Get()->DeleteVariable([](RunContext s) {}, shandle.ctx, var);
+        if (engine_ptr) {
+          engine_ptr->DeleteVariable([](RunContext s) {}, shandle.ctx, var);
+        }
       } else {
         Storage::Handle h = this->shandle;
-        Engine::Get()->DeleteVariable([h](RunContext s) {
-            Storage::Get()->Free(h);
-          }, shandle.ctx, var);
+        if (engine_ptr) {
+          Engine::Get()->DeleteVariable([h](RunContext s) {
+              Storage::Get()->Free(h);
+            }, shandle.ctx, var);
+        }
       }
     }
   };
diff --git a/include/mxnet/resource.h b/include/mxnet/resource.h
index ad6a6f11fd95..6c63d2b4d394 100644
--- a/include/mxnet/resource.h
+++ b/include/mxnet/resource.h
@@ -115,6 +115,11 @@ class ResourceManager {
    * \return Resource manager singleton.
    */
   static ResourceManager *Get();
+
+  /**
+   * Destroy the Resource manager singleton.
+   */
+  static void Shutdown();
 };
 }  // namespace mxnet
 #endif  // MXNET_RESOURCE_H_
diff --git a/make/dato-config.mk b/make/dato-config.mk
index 482d086a50d0..e77b6931355d 100644
--- a/make/dato-config.mk
+++ b/make/dato-config.mk
@@ -26,12 +26,15 @@ export NVCC = ${NVCC}
 
 # whether compile with debug
 DEBUG = 0
+WIN32 = ${WIN32}
 
 # the additional link flags you want to add
-ADD_LDFLAGS = -L${DEPS}/lib -L${DEPS}/lib64 ${SHARED_LINKER_FLAGS}
+ADD_LDFLAGS = ${ADD_LDFLAGS}
+ADD_LDFLAGS += -L${DEPS}/lib -L${DEPS}/lib64 ${SHARED_LINKER_FLAGS}
 
 # the additional compile flags you want to add
-ADD_CFLAGS = -I${DEPS}/include
+ADD_CFLAGS = ${ADD_CFLAGS}
+ADD_CFLAGS += -I${DEPS}/include
 
 #---------------------------------------------
 # matrix computation libraries for CPU/GPU
@@ -58,6 +61,11 @@ USE_OPENCV = 0
 
 # use openmp for parallelization
 USE_OPENMP = ${USE_OPENMP}
+ifneq ($(USE_OPENMP), 1)
+	export NO_OPENMP = 1
+	ADD_CFLAGS += -DDISABLE_OPENMP
+endif
+
 
 # choose the version of blas you want to use
 # can be: mkl, blas, atlas, openblas
diff --git a/mshadow b/mshadow
index 29863c50443d..2068a32beeee 160000
--- a/mshadow
+++ b/mshadow
@@ -1 +1 @@
-Subproject commit 29863c50443d8338cdfcfb5836a4dd854b2b0e75
+Subproject commit 2068a32beeee5ccc7b5a69bccea4f1756dc3bff6
diff --git a/ps-lite b/ps-lite
index ca2a28e27a6d..e86dac79d4ae 160000
--- a/ps-lite
+++ b/ps-lite
@@ -1 +1 @@
-Subproject commit ca2a28e27a6d3b305d14222f5aa44d419a1a8c14
+Subproject commit e86dac79d4ae93274af8abf6c20c91dc118dc9a2
diff --git a/scripts/build.sh b/scripts/build.sh
new file mode 100755
index 000000000000..fd64f18a7541
--- /dev/null
+++ b/scripts/build.sh
@@ -0,0 +1,29 @@
+#!/bin/bash
+
+SCRIPT_DIR=$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )
+WORKSPACE=${SCRIPT_DIR}/..
+
+cd ${WORKSPACE}
+source ${WORKSPACE}/scripts/python_env.sh
+
+if [ -d "/usr/local/cuda-7.5" ]; then
+	./configure --cleanup_if_invalid --yes --cuda_path=/usr/local/cuda-7.5
+	HAS_GPU=1
+else
+	./configure --cleanup_if_invalid --yes
+fi
+
+make clean_all
+make -j4
+
+if [[ $OSTYPE != msys ]]; then
+  nosecmd="${PYTHON_EXECUTABLE} ${NOSETEST_EXECUTABLE}"
+else
+  nosecmd="${NOSETEST_EXECUTABLE}"
+fi
+
+if [ ! -z "$HAS_GPU" ]; then
+  ${nosecmd} -v --with-id ${WORKSPACE}/tests/python/unittest ${WORKSPACE}/tests/python/train ${WORKSPACE}/tests/python/gpu --with-xunit --xunit-file=alltests.nosetests.xml
+else
+  ${nosecmd} -v --with-id ${WORKSPACE}/tests/python/unittest ${WORKSPACE}/tests/python/train --with-xunit --xunit-file=alltests.nosetests.xml
+fi
diff --git a/src/c_api/c_api.cc b/src/c_api/c_api.cc
index 407a1a3696ae..b030f8956373 100644
--- a/src/c_api/c_api.cc
+++ b/src/c_api/c_api.cc
@@ -107,6 +107,8 @@ int MXRandomSeed(int seed) {
 int MXNotifyShutdown() {
   API_BEGIN();
   Engine::Get()->NotifyShutdown();
+  ResourceManager::Shutdown();
+  Engine::Shutdown();
   API_END();
 }
 
diff --git a/src/common/object_pool.h b/src/common/object_pool.h
index ac12e2b30c6e..3717bb1b9cc8 100644
--- a/src/common/object_pool.h
+++ b/src/common/object_pool.h
@@ -52,7 +52,7 @@ class ObjectPool {
    * \brief Internal structure to hold pointers.
    */
   struct LinkedList {
-#if defined(_MSC_VER)
+#if defined(_MSC_VER) || defined(_WIN32)
     T t;
     LinkedList* next{nullptr};
 #else
@@ -168,7 +168,7 @@ void ObjectPool<T>::AllocateChunk() {
   static_assert(alignof(LinkedList) % alignof(T) == 0, "ObjectPooll Invariant");
   static_assert(kPageSize % alignof(LinkedList) == 0, "ObjectPooll Invariant");
   void* new_chunk_ptr;
-#ifdef _MSC_VER
+#if defined(_MSC_VER) || defined(_WIN32)
   new_chunk_ptr = _aligned_malloc(kPageSize, kPageSize);
   CHECK_NE(new_chunk_ptr, NULL) << "Allocation failed";
 #else
diff --git a/src/engine/engine.cc b/src/engine/engine.cc
index ae72861260e1..2067e3b5c44b 100644
--- a/src/engine/engine.cc
+++ b/src/engine/engine.cc
@@ -39,13 +39,29 @@ inline Engine* CreateEngine() {
 }
 }  // namespace engine
 
-std::shared_ptr<Engine> Engine::_GetSharedRef() {
+
+static bool engine_shutdown = false;
+
+std::shared_ptr<Engine>& Engine::_GetSharedRef() {
+  if (engine_shutdown) LOG(FATAL) << "Call GetRef() after engine already shutdown" << std::endl;
   static std::shared_ptr<Engine> sptr(engine::CreateEngine());
   return sptr;
 }
 
 Engine* Engine::Get() {
+  if (engine_shutdown) {
+    return nullptr;
+  }
   static Engine *inst = _GetSharedRef().get();
   return inst;
 }
+
+void Engine::Shutdown() {
+  if (!engine_shutdown) {
+    Engine::Get()->WaitForAll();
+    _GetSharedRef().reset();
+    engine_shutdown = true;
+  }
+}
+
 }  // namespace mxnet
diff --git a/src/engine/threaded_engine_perdevice.cc b/src/engine/threaded_engine_perdevice.cc
index 62738cbaedac..5e981b547435 100644
--- a/src/engine/threaded_engine_perdevice.cc
+++ b/src/engine/threaded_engine_perdevice.cc
@@ -38,9 +38,10 @@ class ThreadedEnginePerDevice : public ThreadedEngine {
     // create CPU task
     int cpu_priority_nthreads = dmlc::GetEnv("MXNET_CPU_PRIORITY_NTHREADS", 4);
     cpu_priority_worker_.reset(new ThreadWorkerBlock<kPriorityQueue>());
+    auto blockptr = cpu_priority_worker_.get();
     cpu_priority_worker_->pool.reset(new ThreadPool(
-        cpu_priority_nthreads, [this] {
-          this->CPUWorker(cpu_priority_worker_.get());
+        cpu_priority_nthreads, [this, blockptr] {
+          this->CPUWorker(blockptr);
         }));
     // GPU tasks will be created lazily
   }
diff --git a/src/resource.cc b/src/resource.cc
index b791960fa812..87094f9e7f2d 100644
--- a/src/resource.cc
+++ b/src/resource.cc
@@ -190,8 +190,22 @@ class ResourceManagerImpl : public ResourceManager {
 };
 }  // namespace resource
 
+static resource::ResourceManagerImpl* instance_ptr = nullptr;
+static bool rm_shutdown = false;
+
 ResourceManager* ResourceManager::Get() {
-  static resource::ResourceManagerImpl inst;
-  return &inst;
+  if (instance_ptr == nullptr) {
+    if (rm_shutdown) LOG(FATAL) << "Resource manager already shutdone" << std::endl;
+    instance_ptr = new resource::ResourceManagerImpl();
+  }
+  return instance_ptr;
+}
+
+void ResourceManager::Shutdown() {
+  if (instance_ptr != nullptr) {
+    delete instance_ptr;
+    instance_ptr = nullptr;
+    rm_shutdown = true;
+  }
 }
 }  // namespace mxnet
diff --git a/src/storage/cpu_device_storage.h b/src/storage/cpu_device_storage.h
index 6838af037535..bfe364302528 100644
--- a/src/storage/cpu_device_storage.h
+++ b/src/storage/cpu_device_storage.h
@@ -38,7 +38,7 @@ class CPUDeviceStorage {
 };  // class CPUDeviceStorage
 
 inline void* CPUDeviceStorage::Alloc(size_t size) {
-#if _MSC_VER
+#if defined(_MSC_VER) || defined(_WIN32)
   void* ptr;
   ptr = _aligned_malloc(size, alignment_);
   return CHECK_NOTNULL(ptr);
@@ -51,7 +51,7 @@ inline void* CPUDeviceStorage::Alloc(size_t size) {
 }
 
 inline void CPUDeviceStorage::Free(void* ptr) {
-#if _MSC_VER
+#if defined(_MSC_VER) || defined(_WIN32)
   _aligned_free(ptr);
 #else
   free(ptr);
diff --git a/src/symbol/graph_executor.cc b/src/symbol/graph_executor.cc
index c00691bb6daf..c27b2244c7bd 100644
--- a/src/symbol/graph_executor.cc
+++ b/src/symbol/graph_executor.cc
@@ -261,10 +261,13 @@ GraphExecutor::GetOpExecEntry(uint32_t nid) {
 }
 
 GraphExecutor::~GraphExecutor() {
-  Engine::Get()->WaitForAll();
-  // need to delete the operators before delete the NDArray they referenced.
-  for (OpNode& node : op_nodes_) {
-    node.DeleteOperator();
+  auto engine_ptr = Engine::Get();
+  if (engine_ptr) { 
+    engine_ptr->WaitForAll();
+    // need to delete the operators before delete the NDArray they referenced.
+    for (OpNode& node : op_nodes_) {
+      node.DeleteOperator();
+    }
   }
 }
 
diff --git a/tests/python/common/get_data.py b/tests/python/common/get_data.py
index e880a9ab0b8d..4a3afd9f1f19 100644
--- a/tests/python/common/get_data.py
+++ b/tests/python/common/get_data.py
@@ -15,14 +15,14 @@ def download_file(url, target_file):
 # download mnist.pkl.gz
 def GetMNIST_pkl():
     if not os.path.isdir("data/"):
-        os.system("mkdir data/")
+        os.mkdir("data")
     if not os.path.exists('data/mnist.pkl.gz'):
         download_file("http://deeplearning.net/data/mnist/mnist.pkl.gz", "data/mnist.pkl.gz")
 
 # download ubyte version of mnist and untar
 def GetMNIST_ubyte():
     if not os.path.isdir("data/"):
-        os.system("mkdir data/")
+        os.mkdir("data")
     if (not os.path.exists('data/train-images-idx3-ubyte')) or \
        (not os.path.exists('data/train-labels-idx1-ubyte')) or \
        (not os.path.exists('data/t10k-images-idx3-ubyte')) or \
@@ -36,7 +36,7 @@ def GetMNIST_ubyte():
 # download cifar
 def GetCifar10():
     if not os.path.isdir("data/"):
-        os.system("mkdir data/")
+        os.mkdir("data")
     if not os.path.exists('data/cifar10.zip'):
         download_file("http://webdocs.cs.ualberta.ca/~bx3/data/cifar10.zip", "data/cfar10.zip")
         os.chdir("./data")