@@ -73,6 +73,7 @@ All the lists in this page are either in alphabetical order or chronological ord
 
 - [C++ Concurrency in Action: Practical Multithreading](https://www.manning.com/books/c-plus-plus-concurrency-in-action) - Anthony Williams 2012
 - [The Art of Multiprocessor Programming](https://www.amazon.com/Art-Multiprocessor-Programming-Revised-Reprint/dp/0123973376/ref=sr_1_1?ie=UTF8&qid=1438003865&sr=8-1&keywords=maurice+herlihy) - Maurice Herlihy 2012
+- [Parallel Computing: Theory and Practice](http://www.cs.cmu.edu/afs/cs/academic/class/15210-f15/www/tapp.html#ch:work-stealing) - Umut A. Acar 2016
 
 #### Tutorials
 
 
@@ -108,8 +108,6 @@ void run_tbb(MNIST& D, unsigned num_threads) {
     }
   } // End of all epoch
 
-  return;
-
   for(size_t i=0; i<num_par_shf; i++) {
     shuffle_tasks[i]->try_put(continue_msg());
   }
 
@@ -75,6 +75,7 @@ class Executor {
   struct Worker {
     std::mt19937 rdgen { std::random_device{}() };
     WorkStealingQueue<Node*> queue;
+    std::optional<Node*> cache;
   };
 
   struct PerThread {
@@ -226,7 +227,7 @@ class Executor {
     void _spawn(unsigned);
     void _exploit_task(unsigned, std::optional<Node*>&);
     void _explore_task(unsigned, std::optional<Node*>&);
-    void _schedule(Node*);
+    void _schedule(Node*, bool);
     void _schedule(PassiveVector<Node*>&);
     void _invoke(unsigned, Node*);
     void _invoke_static_work(unsigned, Node*);
@@ -417,11 +418,22 @@ inline void Executor::_exploit_task(unsigned i, std::optional<Node*>& t) {
 
     do {
       _invoke(i, *t);
-      t = worker.queue.pop();
+
+      if(worker.cache) {
+        t = *worker.cache;
+        worker.cache = std::nullopt;
+      }
+      else {
+        t = worker.queue.pop();
+      }
+
     } while(t);
 
     --_num_actives;
   }
+  else {
+    assert(!_workers[i].cache);
+  }
 }
 
 // Function: _wait_for_tasks
@@ -478,7 +490,7 @@ inline void Executor::remove_observer() {
 // Procedure: _schedule
 // The main procedure to schedule a give task node.
 // Each task node has two types of tasks - regular and subflow.
-inline void Executor::_schedule(Node* node) {
+inline void Executor::_schedule(Node* node, bool bypass) {
 
   // module node need another initialization
   if(node->_module != nullptr && !node->is_spawned()) {
@@ -493,7 +505,13 @@ inline void Executor::_schedule(Node* node) {
 
   // caller is a worker to this pool
   if(auto& pt = _per_thread(); pt.pool == this) {
-    _workers[pt.worker_id].queue.push(node);
+    if(!bypass) {
+      _workers[pt.worker_id].queue.push(node);
+    }
+    else {
+      assert(!_workers[pt.worker_id].cache);
+      _workers[pt.worker_id].cache = node;
+    }
     return;
   }
 
@@ -673,12 +691,22 @@ inline void Executor::_invoke(unsigned me, Node* node) {
   }
 
   // At this point, the node storage might be destructed.
+  Node* cache {nullptr};
+
   for(size_t i=0; i<num_successors; ++i) {
     if(--(node->_successors[i]->_num_dependents) == 0) {
-      _schedule(node->_successors[i]);
+      if(cache) {
+        //_schedule(node->_successors[i]);
+        _schedule(cache, false);
+      }
+      cache = node->_successors[i];
     }
   }
 
+  if(cache) {
+    _schedule(cache, true);
+  }
+
   // A node without any successor should check the termination of topology
   if(num_successors == 0) {
     if(--(node->_topology->_num_sinks) == 0) {
Original file line number	Diff line number	Diff line change
@@ -108,8 +108,6 @@ void run_tbb(MNIST& D, unsigned num_threads) {
108	108	}
109	109	} // End of all epoch
110	110
111		- return;
112		-
113	111	for(size_t i=0; i<num_par_shf; i++) {
114	112	shuffle_tasks[i]->try_put(continue_msg());
115	113	}