diff --git a/core/engine/cluster/cloud/cluster.sh b/core/engine/cluster/cloud/cluster.sh
index 8f8c5479df508dfc5e74ee936b665ba08d4647b1..399a21e78aa2eba2489c8aa0b4f2214328bd0a50 100644
--- a/core/engine/cluster/cloud/cluster.sh
+++ b/core/engine/cluster/cloud/cluster.sh
@@ -59,6 +59,7 @@ function _gen_mpi_config() {
       -e "s#<$ OUTPUT_PATH $>#$OUTPUT_PATH#g" \
       -e "s#<$ THIRDPARTY_PATH $>#$THIRDPARTY_PATH#g" \
       -e "s#<$ CPU_NUM $>#$max_thread_num#g" \
+      -e "s#<$ USE_PYTHON3 $>#$USE_PYTHON3#g" \
       -e "s#<$ FLAGS_communicator_is_sgd_optimizer $>#$FLAGS_communicator_is_sgd_optimizer#g" \
       -e "s#<$ FLAGS_communicator_send_queue_size $>#$FLAGS_communicator_send_queue_size#g" \
       -e "s#<$ FLAGS_communicator_thread_pool_size $>#$FLAGS_communicator_thread_pool_size#g" \
@@ -76,6 +77,7 @@ function _gen_k8s_config() {
       -e "s#<$ AFS_REMOTE_MOUNT_POINT $>#$AFS_REMOTE_MOUNT_POINT#g" \
       -e "s#<$ OUTPUT_PATH $>#$OUTPUT_PATH#g" \
       -e "s#<$ CPU_NUM $>#$max_thread_num#g" \
+      -e "s#<$ USE_PYTHON3 $>#$USE_PYTHON3#g" \
       -e "s#<$ FLAGS_communicator_is_sgd_optimizer $>#$FLAGS_communicator_is_sgd_optimizer#g" \
       -e "s#<$ FLAGS_communicator_send_queue_size $>#$FLAGS_communicator_send_queue_size#g" \
       -e "s#<$ FLAGS_communicator_thread_pool_size $>#$FLAGS_communicator_thread_pool_size#g" \
diff --git a/core/engine/cluster/cloud/k8s_config.ini.template b/core/engine/cluster/cloud/k8s_config.ini.template
index 904bfbc5e1453f90ec1163d1681d554b52dae45f..8979cc6f0d996c132fbc2259a01134ba4a8a1ee5 100644
--- a/core/engine/cluster/cloud/k8s_config.ini.template
+++ b/core/engine/cluster/cloud/k8s_config.ini.template
@@ -19,6 +19,7 @@ afs_local_mount_point="/root/paddlejob/workspace/env_run/afs/"
 # 新k8s afs挂载帮助文档: http://wiki.baidu.com/pages/viewpage.action?pageId=906443193
 
 PADDLE_PADDLEREC_ROLE=WORKER
+use_python3=<$ USE_PYTHON3 $>
 CPU_NUM=<$ CPU_NUM $>
 GLOG_v=0
 
diff --git a/core/engine/cluster/cloud/mpi_config.ini.template b/core/engine/cluster/cloud/mpi_config.ini.template
index 8312d46a01449b3d6eac322b098d5b029bb67f86..7d9f7fbb97a53c23e566e925a87eae990cef9f2a 100644
--- a/core/engine/cluster/cloud/mpi_config.ini.template
+++ b/core/engine/cluster/cloud/mpi_config.ini.template
@@ -17,6 +17,7 @@ output_path=<$ OUTPUT_PATH $>
 thirdparty_path=<$ THIRDPARTY_PATH $>
 
 PADDLE_PADDLEREC_ROLE=WORKER
+use_python3=<$ USE_PYTHON3 $>
 CPU_NUM=<$ CPU_NUM $>
 GLOG_v=0
 
diff --git a/core/engine/cluster/cluster.py b/core/engine/cluster/cluster.py
index 7dbb5708e572340c37265972e541bb00ef2ee195..c94124581fcf3a591a924211ae36ce87bae97586 100644
--- a/core/engine/cluster/cluster.py
+++ b/core/engine/cluster/cluster.py
@@ -159,23 +159,28 @@ class ClusterEnvBase(object):
         self.cluster_env["PADDLE_VERSION"] = self.backend_env.get(
             "config.paddle_version", "1.7.2")
 
+        # python_version
+        self.cluster_env["USE_PYTHON3"] = self.backend_env.get(
+            "config.use_python3", "0")
+
         # communicator
+        max_thread_num = int(envs.get_runtime_environ("max_thread_num"))
         self.cluster_env[
             "FLAGS_communicator_is_sgd_optimizer"] = self.backend_env.get(
                 "config.communicator.FLAGS_communicator_is_sgd_optimizer", 0)
         self.cluster_env[
             "FLAGS_communicator_send_queue_size"] = self.backend_env.get(
-                "config.communicator.FLAGS_communicator_send_queue_size", 5)
+                "config.communicator.FLAGS_communicator_send_queue_size", max_thread_num)
         self.cluster_env[
             "FLAGS_communicator_thread_pool_size"] = self.backend_env.get(
                 "config.communicator.FLAGS_communicator_thread_pool_size", 32)
         self.cluster_env[
             "FLAGS_communicator_max_merge_var_num"] = self.backend_env.get(
-                "config.communicator.FLAGS_communicator_max_merge_var_num", 5)
+                "config.communicator.FLAGS_communicator_max_merge_var_num", max_thread_num)
         self.cluster_env[
             "FLAGS_communicator_max_send_grad_num_before_recv"] = self.backend_env.get(
                 "config.communicator.FLAGS_communicator_max_send_grad_num_before_recv",
-                5)
+                max_thread_num)
         self.cluster_env["FLAGS_communicator_fake_rpc"] = self.backend_env.get(
             "config.communicator.FLAGS_communicator_fake_rpc", 0)
         self.cluster_env["FLAGS_rpc_retry_times"] = self.backend_env.get(
diff --git a/doc/distributed_train.md b/doc/distributed_train.md
index 9e7dbf1bd903e459d78f18f66e5893cb3d3ced1b..9c07babaa52898c2dd5ea0cace4e0aec6f68a130 100644
--- a/doc/distributed_train.md
+++ b/doc/distributed_train.md
@@ -69,6 +69,12 @@ dataset:
   data_path: "{workspace}/data/sample_data/train"
   sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
   dense_slots: "dense_var:13"
+
+phase:
+- name: phase1
+  model: "{workspace}/model.py"
+  dataset_name: dataloader_train 
+  thread_num: 1
 ```
 
 分布式的训练配置可以改为：
@@ -101,6 +107,13 @@ dataset:
   data_path: "{workspace}/train_data"
   sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
   dense_slots: "dense_var:13"
+
+phase:
+- name: phase1
+  model: "{workspace}/model.py"
+  dataset_name: dataloader_train 
+  # 分布式训练节点的CPU_NUM环境变量与thread_num相等，多个phase时，取最大的thread_num
+  thread_num: 1
 ```
 
 除此之外，还需关注数据及模型加载的路径，一般而言：
@@ -120,6 +133,8 @@ cluster_type: mpi # k8s 可选
 config:
   # 填写任务运行的paddle官方版本号 >= 1.7.2， 默认1.7.2
   paddle_version: "1.7.2" 
+  # 是否使用PaddleCloud运行环境下的Python3，默认使用python2
+  use_python3: 1
 
   # hdfs/afs的配置信息填写
   fs_name: "afs://xxx.com"
@@ -140,11 +155,13 @@ config:
 
   # paddle参数服务器分布式底层超参，无特殊需求不理不改
   communicator:
+    # 使用SGD优化器时，建议设置为1
     FLAGS_communicator_is_sgd_optimizer: 0
+    # 以下三个变量默认都等于训练时的线程数：CPU_NUM
     FLAGS_communicator_send_queue_size: 5
-    FLAGS_communicator_thread_pool_size: 32
     FLAGS_communicator_max_merge_var_num: 5
     FLAGS_communicator_max_send_grad_num_before_recv: 5
+    FLAGS_communicator_thread_pool_size: 32
     FLAGS_communicator_fake_rpc: 0
     FLAGS_rpc_retry_times: 3
   
@@ -232,7 +249,7 @@ phase:
 再新增`backend.yaml`
 ```yaml
 backend: "PaddleCloud"
-cluster_type: mpi 
+cluster_type: mpi # k8s可选
 
 config:
   paddle_version: "1.7.2" 
@@ -317,7 +334,7 @@ phase:
 
 ```yaml
 backend: "PaddleCloud"
-cluster_type: k8s # k8s 可选
+cluster_type: k8s # mpi 可选
 
 config:
   # 填写任务运行的paddle官方版本号 >= 1.7.2， 默认1.7.2
@@ -399,7 +416,7 @@ phase:
 再新增`backend.yaml`
 ```yaml
 backend: "PaddleCloud"
-cluster_type: k8s # k8s 可选
+cluster_type: k8s # mpi 可选
 
 config:
   # 填写任务运行的paddle官方版本号 >= 1.7.2， 默认1.7.2
diff --git a/doc/yaml.md b/doc/yaml.md
index 4d08ef72253b5fa1371faa5890c53ba32c0ca2e6..0450e50b6866f84a5295600bdcd2da94535ce4ba 100644
--- a/doc/yaml.md
+++ b/doc/yaml.md
@@ -1,4 +1,4 @@
-# PaddleRec yaml配置说明
+# PaddleRec config.yaml配置说明
 
 ## 全局变量
 
@@ -70,3 +70,20 @@
 | optimizer.learning_rate | float  |       > 0        |    否    |         指定学习率          |
 |           reg           | float  |       > 0        |    否    | L2正则化参数，只在SGD下生效 |
 |         others          |   /    |        /         |    /     |   由各个模型组网独立指定    |
+
+
+# PaddleRec backend.yaml配置说明
+
+## 全局变量
+
+ |   名称    |         类型          |                         取值                          | 是否必须 |                      作用描述                      |
+ | :-------: | :-------------------: | :---------------------------------------------------: | :------: | :------------------------------------------------: |
+ | backend |        string         |      paddlecloud/k8s       |    是    |           使用PaddleCloud平台提交，还是在公有云K8S集群提交            |
+ |   cluster_type    | string  | mpi/k8s |    是    |             指定运行的计算集群： mpi 还是 k8s             |
+
+ ## config
+ |             名称              |     类型     |                     取值                      | 是否必须 |                               作用描述                               |
+| :---------------------------: | :----------: | :-------------------------------------------: | :------: | :------------------------------------------------------------------: |
+|             paddle_version             |    string    | paddle官方版本号，如1.7.2/1.8.0/1.8.3等 |    否    |           指定运行训练使用的Paddle版本，默认1.7.2            |
+|            use_python3             |    int    |                0（默认）/1                |    否    |              指定是否使用python3进行训练                           |
+|          fs_name           |    string    |         "afs://xxx.com"         |    是    |                            hadoop集群名称所需配置                            |
\ No newline at end of file