diff --git a/doc/README.md b/doc/README.md
index d16d21d6..bb6a624e 100644
--- a/doc/README.md
+++ b/doc/README.md
@@ -27,7 +27,7 @@ Sample configuration files can be found at example folder
 * [AlexNet for ImageNet](../example/ImageNet)
 
 ### Parameter Setting
-To understand more about cxxnet setting, this section will introduce all parameters of cxxnet and the effect of parameters. In general, cxxnet configuration file contains 4 kinds of configurations in a single file:
+To understand more about cxxnet setting, this section will introduce all parameters of cxxnet and the effect of parameters. In general, cxxnet configuration file contains 3 kinds of configurations in a single file:
 
 * [Data Input Iterator Setting](io.md)
   - Set input data configurations.
diff --git a/doc/global.md b/doc/global.md
new file mode 100644
index 00000000..5a38d79b
--- /dev/null
+++ b/doc/global.md
@@ -0,0 +1,127 @@
+#### Introduction
+This page will introduce global setting in cxxnet, including:
+* [Device Selection](#set-working-hardware)
+* [Printing Control](#print-information)
+* [Training Round](#set-round-of-training)
+* [Saving Model and Continue Training](#saving-model-and-continue-training)
+* [Prediction](#prediction)
+* [Extract Feature](#extract-feature)
+* [Fine-tune](#fine-tune)
+
+
+#### Global setting
+Global setting is to set parameters which are used globally. Global setting parameters are in any sections outside netconfig and iter. To understand better, you may find compare to the [MNIST.conf](example/MNIST.conf) can be helpful.
+
+Indeed you can set any parameter in the global area. Local setting has higher privilege and is able to override the global setting if they are in same name.
+
+
+#### Set Global Setting via Command Line
+* Besides setting parameters via config file, cxxnet also support set parameters via command line. The syntax is like:
+```bash
+cxxnet_learner config.conf dev=gpu
+```
+Then the program will run in gpu mode. Note in command line mode, we must ensure ```dev=gpu``` contains no space so that they are passed in as single argument. The settings in command line will override settings in config file; this allows easy change of parameters via command line if you want to quick try different kinds of options.
+
+
+#### Set working hardware
+* To use CPU, set the field
+```bash
+dev = cpu
+```
+* To use GPU, set the field
+```bash
+dev = gpu
+```
+We can also set specific device (say device 1) by using
+```bash
+dev = gpu:1
+```
+* To use multi-GPU, set the field with the corresponding device id
+```bash
+dev = gpu:0,1,2,3
+```
+In default, it is `dev=gpu`
+
+
+#### Print information
+* To print training error evaluation, just set this field to 1
+```bash
+eval_train = 1
+```
+* in default this field is 0, which means cxxnet won't print anything about training error.
+* To turn off all information while training, set this field to 1
+```bash
+silent = 1
+```
+* In default this field is 0
+* To control print frequent, change this field
+```bash
+print_step = 100
+```
+* In default it will print every 100 batch
+
+
+#### Set round of training
+There are two field handle training round together: _**num_round**_ and _**max_round**_
+* _**num_round**_ is used for number of round to train
+* _**max_round**_ is used for maximum number of round to train from now on
+```bash
+num_round = 15
+max_round = 15
+```
+This configuration will make cxxnet train for 15 rounds on the training data.
+
+More examples,
+```bash
+num_round = 50
+max_round = 2
+```
+If we have a model trained 40 rounds, then use this configuration continue to train, cxxnet will stop at the 42 round.
+
+
+#### Saving model and continue training
+* To save model while training round, set this field to saving frequent(a number)
+```bash
+save_model = 2
+model_dir = path_of_dir_to_save_model
+```
+* In default, this field is 1, means cxxnet will save a model in every round
+* To continue a training process, you need to set model_in as the input snapshot you want to continue from
+```conf
+model_in = path of model file
+```
+* Alternatively, if you save model every round (save_model=1), then you can use option continue, cxxnet will automatically search the latest model and start from that model
+```conf
+continue = 1
+```
+In default, if neither of the two values is set, cxxnet will start training from start.
+
+
+#### Prediction
+* In default, cxxnet treats the configuration file as a training configuration. To make it predict, you need to add extra data iterator and specify the task to be `pred` and model you want to use to do prediction. For example
+```bash
+# Data Iterator Setting
+pred = pred.txt
+iter = mnist
+iterator_optition_1 = ..
+iterator_optition_1 = ...
+iter = end
+# Global Setting
+task = pred
+model_in = ./models/0014.model
+```
+* In which the _*mode_in*_ is the path to the model which we need to use for prediction. The _*pred*_ field is the file we will save the result. The iterator configuration is same to traditional iterator.
+* To get the softmax prediction directly, set the task to
+```bash
+task = pred_raw
+```
+#### Extract Feature
+* To extract feature, you need to set task to ```extract_feature```with node name
+```bash
+task = extract_feature
+extract_node_name = 45
+```
+If you didn't set name for layer, just use the number as name; If you set special name, use the name. For convenient, a special name ```top``` is used for extract topest layer behind loss layer.
+
+#### Fine-tune
+TODO