diff --git a/pgl/layers/graph_pool.py b/pgl/layers/graph_pool.py
index a88468f7b6ef12131c4554db9339c35472a66f11..a1b6249c21d9022a30e8f51f2a6950a187ee3b14 100644
--- a/pgl/layers/graph_pool.py
+++ b/pgl/layers/graph_pool.py
@@ -19,7 +19,7 @@ from pgl import graph_wrapper
 from pgl.utils import paddle_helper
 from pgl.utils import op
 
-__all__ = ['graph_pooling']
+__all__ = ['graph_pooling', 'graph_norm']
 
 
 def graph_pooling(gw, node_feat, pool_type):
@@ -40,3 +40,29 @@ def graph_pooling(gw, node_feat, pool_type):
     graph_feat = op.nested_lod_reset(node_feat, gw.graph_lod)
     graph_feat = fluid.layers.sequence_pool(graph_feat, pool_type)
     return graph_feat
+
+
+def graph_norm(gw, feature):
+    """Implementation of graph normalization
+   
+    Reference Paper: BENCHMARKING GRAPH NEURAL NETWORKS
+   
+    Each node features is divied by sqrt(num_nodes) per graphs.
+
+    Args:
+        gw: Graph wrapper object (:code:`StaticGraphWrapper` or :code:`GraphWrapper`)
+
+        graph_level (default: False):  If :code:`graph_level=True` return shape (num_graphs, 1)
+                    elif :code:`graph_level=False return shape (num_nodes, 1)`
+
+    Return:
+        A tensor with shape (num_graphs, 1) or (num_node, 1)
+    """
+    nodes = fluid.layers.fill_constant(
+        [gw.num_nodes, 1], dtype="float32", value=1.0)
+    norm = graph_pooling(gw, nodes, pool_type="sum")
+    norm = fluid.layers.sqrt(norm)
+    feature_lod = op.nested_lod_reset(feature, gw.graph_lod)
+    norm = fluid.layers.sequence_expand_as(norm, feature_lod)
+    norm.stop_gradient = True
+    return feature_lod / norm