diff --git a/examples/breakout_impala_binding_test.yaml b/examples/breakout_impala_binding_test.yaml
new file mode 100644
index 0000000..cdd4acf
--- /dev/null
+++ b/examples/breakout_impala_binding_test.yaml
@@ -0,0 +1,37 @@
+alg_para:
+  alg_name: IMPALAOpt
+  alg_config: 
+    train_per_checkpoint: 1
+    prepare_times_per_train: 1
+    BATCH_SIZE: 512
+
+
+env_para:
+  env_name: VectorAtariEnv
+  env_info: 
+    name: BreakoutNoFrameskip-v4
+    vision: False
+    vector_env_size: 1
+
+agent_para:
+  agent_name: AtariImpalaOpt
+  agent_num : 1
+  agent_config:
+    max_steps: 128
+    complete_step: 10020000
+
+model_para:
+  actor:
+    model_name: ImpalaCnnOpt
+    state_dim: [84,84,4]
+    input_dtype: uint8
+    state_mean: 0.0
+    state_std: 255.0
+    action_dim: 4
+    model_config:
+      LR: 0.0005
+      sample_batch_step: 128
+      grad_norm_clip: 40.0
+
+env_num: 32
+speedup: 2
diff --git a/xt/framework/broker.py b/xt/framework/broker.py
index 9fa1605..2a205df 100644
--- a/xt/framework/broker.py
+++ b/xt/framework/broker.py
@@ -496,10 +496,44 @@ def create_explorer(self, config_info):
         p = Process(target=explorer.start)
         p.start()
 
-        cpu_count = psutil.cpu_count()
-        if speedup and cpu_count > (env_num + start_core):
+        # updated balanced core binding
+        """
+        Speed up training by setting 'speedup' through config.yaml
+        0: System scheduling without core binding
+        1: [default] Binding sequentially according to the core id and explore id
+        2: Balanced core binding.
+        3: Pipeline (to be updated)
+        """
+        # cpu_count = psutil.cpu_count()
+        # if speedup and cpu_count > (env_num + start_core):
+        #     _p = psutil.Process(p.pid)
+        #     _p.cpu_affinity([start_core + env_id])
+
+        socket = 0
+        with open('/proc/cpuinfo', 'r') as f:
+            for line in f:  # number of sockets = maximum physical cpu id
+                if line.lower().startswith('physical'):  # physical id : xxx
+                    socket = max(socket, int(line.split()[-1]))
+        socket = socket + 1
+        logical, physical = psutil.cpu_count(), psutil.cpu_count(logical=False)
+
+        if speedup == 3: # pipeline: to be updated
+            pass
+        elif speedup == 2 and logical >= env_num:  # balanced core binding
+            base = ((env_id // physical) % (logical // physical)) * physical
+            env = env_id % physical
+            cpu = env // socket + (env % socket) * (physical // socket) + base
+            _p = psutil.Process(p.pid)
+            _p.cpu_affinity([cpu])
+            logging.info("speedup({:d}): binding env {:d} to CPU {:d}".format(speedup, env_id, cpu))
+        elif speedup == 1 and logical > (env_num + start_core):  # sequential core binding
             _p = psutil.Process(p.pid)
             _p.cpu_affinity([start_core + env_id])
+            logging.info("speedup({:d}): binding env {:d} to CPU {:d}"
+                         .format(speedup, env_id, start_core + env_id))
+        else:  # scheduled by os
+            logging.info("speedup({:d}): no binding".format(speedup))
+
 
         self.send_explorer_q.update({env_id: send_explorer})
         self.explore_process.update({env_id: p})