Add XGBoost demo

demos/README.md

@@ -7,3 +7,4 @@ This directory contains sample projects that demonstrate how Occlum can be used
 * `hello_bazel/`: A sample C++ project built with [Bazel](https://bazel.build).
 * `https_server/`: A HTTPS file server based on [Mongoose Embedded Web Server Library](https://github.com/cesanta/mongoose).
 * `tensorflow_lite/`: A demo and benchmark of [Tensorflow Lite](https://www.tensorflow.org/lite) inference engine.
+* `xgboost/`: A demo of [XGBoost](https://xgboost.readthedocs.io/en/latest/).

demos/xgboost/.gitignore

@@ -0,0 +1,5 @@
+xgboost_src/
+data/
+host/
+occlum_workspace/
+tmp_*

demos/xgboost/Makefile

@@ -0,0 +1,48 @@
+CUR_DIR := $(shell dirname $(realpath $(lastword $(MAKEFILE_LIST))))
+XGBOOST_TRACKER := $(realpath $(CUR_DIR)/xgboost_src/dmlc-core/tracker/)
+BIN := xgboost
+DATA := data
+
+.PHONY: all test test-native test-local-cluster clean
+
+all: occlum_workspace
+
+$(BIN):
+	cp xgboost_src/$(BIN) .
+
+$(DATA):
+	rm -rf $(DATA) && mkdir -p $(DATA)
+	cp xgboost_src/demo/binary_classification/agaricus.txt.test $(DATA)
+	cp xgboost_src/demo/binary_classification/agaricus.txt.train $(DATA)
+	cp xgboost_src/demo/binary_classification/mushroom.conf $(DATA)
+	sed -i 's/agaricus/.\/data\/agaricus/g' $(DATA)/mushroom.conf
+	sed -i 's/num_round = 2/num_round = 10/g' $(DATA)/mushroom.conf
+	sed -i '$$a\model_out = "./host/result.model"' $(DATA)/mushroom.conf
+	sed -i '$$a\verbosity = 2' $(DATA)/mushroom.conf
+	sed -i '$$a\nthread = 2' $(DATA)/mushroom.conf
+
+occlum_workspace: $(BIN) $(DATA)
+	rm -rf occlum_workspace && mkdir occlum_workspace
+	cd occlum_workspace && \
+	occlum init && \
+	jq '.process.default_mmap_size = "96MB"' Occlum.json > temp_Occlum.json && \
+	mv temp_Occlum.json Occlum.json && \
+	cp ../$(BIN) image/bin && \
+	cp ../xgboost_src/lib/libxgboost.so image/lib && \
+	mkdir -p image/$(DATA) && \
+	cp -r ../$(DATA)/* image/$(DATA) && \
+	occlum build
+
+test: occlum_workspace
+	@./run_xgboost.sh
+
+test-native: $(BIN) $(DATA)
+	rm -rf host
+	mkdir host
+	./$(BIN) ./$(DATA)/mushroom.conf
+
+test-local-cluster: occlum_workspace
+	@$(XGBOOST_TRACKER)/dmlc-submit --cluster=local --num-workers=2 --log-level=DEBUG ./run_xgboost.sh
+
+clean:
+	rm -rf $(BIN) $(DATA) tmp_* host occlum_workspace

demos/xgboost/README.md

@@ -0,0 +1,22 @@
+# Use XGBoost in SGX with Occlum
+
+Step 1: Download XGBoost and its dependencies, and then build XGBoost
+```
+./build_xgboost.sh
+```
+When completed, the resulting XGBoost can be found in `xgboost_src` directory.
+
+Step 2: To train data with XGBoost in a single process, run
+```
+make test
+```
+
+Step 3: To train data with a two-node XGBoost cluster, run
+```
+make test-local-cluster
+```
+
+Step 4 (Optional): To train data with XGBoost in a single process in Linux, run
+```
+make test-native
+```

demos/xgboost/build_xgboost.sh

@@ -0,0 +1,42 @@
+#!/bin/bash
+set -e
+
+# Install the dependencies
+apt-get update
+apt-get install -y python3.5
+apt-get install -y python3-pip
+apt-get install -y python3-setuptools
+apt-get install -y python-pip
+apt-get install -y python-setuptools
+pip3 install kubernetes
+pip install kubernetes
+
+# Download and build XGBoost
+rm -rf xgboost_src && mkdir xgboost_src
+pushd xgboost_src
+git clone https://github.com/dmlc/xgboost .
+git checkout 6d5b34d82486cd1d0480c548f5d1953834659bd6
+git submodule init
+git submodule update
+git apply ../patch/xgboost-01.diff
+pushd rabit
+git apply ../../patch/rabit-01.diff
+popd
+pushd dmlc-core
+git apply ../../patch/dmlc-core-01.diff
+popd
+mkdir build
+cd build
+cmake ../ \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_C_COMPILER=occlum-gcc -DCMAKE_CXX_COMPILER=occlum-g++
+make -j
+popd
+
+# Prepare data
+pushd xgboost_src/demo/binary_classification
+python mapfeat.py
+python mknfold.py agaricus.txt 1
+popd
+
+echo "Build XGBoost Success!"

demos/xgboost/patch/dmlc-core-01.diff

@@ -0,0 +1,18 @@
+diff --git a/tracker/dmlc_tracker/local.py b/tracker/dmlc_tracker/local.py
+index dff7c17..fca0f9d 100644
+--- a/tracker/dmlc_tracker/local.py
++++ b/tracker/dmlc_tracker/local.py
+@@ -24,6 +24,13 @@ def exec_cmd(cmd, role, taskid, pass_env):
+ 
+     num_retry = env.get('DMLC_NUM_ATTEMPT', 0)
+ 
++    cmd_str = ''
++    for k, v in env.items():
++        if str(k)[0:4] == 'DMLC':
++            strenv = str(k) + '=' + str(v);
++            cmd_str = cmd_str + ' ' + strenv
++    cmd = cmd + cmd_str
++
+     while True:
+         if os.name == 'nt':
+             ret = subprocess.call(cmd, shell=True, env=env)

demos/xgboost/patch/rabit-01.diff

@@ -0,0 +1,36 @@
+diff --git a/src/allreduce_base.cc b/src/allreduce_base.cc
+index 143db6e..a6daf20 100644
+--- a/src/allreduce_base.cc
++++ b/src/allreduce_base.cc
+@@ -486,12 +486,13 @@ AllreduceBase::TryAllreduceTree(void *sendrecvbuf_,
+     // select must return
+     watcher.Poll();
+     // exception handling
+-    for (int i = 0; i < nlink; ++i) {
++    //FIXME:workaround for Occlum
++    /*for (int i = 0; i < nlink; ++i) {
+       // recive OOB message from some link
+       if (watcher.CheckExcept(links[i].sock)) {
+         return ReportError(&links[i], kGetExcept);
+       }
+-    }
++    }*/
+     // read data from childs
+     for (int i = 0; i < nlink; ++i) {
+       if (i != parent_index && watcher.CheckRead(links[i].sock)) {
+@@ -641,12 +642,13 @@ AllreduceBase::TryBroadcast(void *sendrecvbuf_, size_t total_size, int root) {
+     // select
+     watcher.Poll();
+     // exception handling
+-    for (int i = 0; i < nlink; ++i) {
++    //FIXME:workaround for Occlum
++    /*for (int i = 0; i < nlink; ++i) {
+       // recive OOB message from some link
+       if (watcher.CheckExcept(links[i].sock)) {
+         return ReportError(&links[i], kGetExcept);
+       }
+-    }
++    }*/
+     if (in_link == -2) {
+       // probe in-link
+       for (int i = 0; i < nlink; ++i) {

demos/xgboost/patch/xgboost-01.diff

@@ -0,0 +1,21 @@
+diff --git a/src/cli_main.cc b/src/cli_main.cc
+index faa93ad..a201615 100644
+--- a/src/cli_main.cc
++++ b/src/cli_main.cc
+@@ -339,6 +339,16 @@ int CLIRunTask(int argc, char *argv[]) {
+     printf("Usage: <config>\n");
+     return 0;
+   }
++
++  //FIXME:workaroud for local distributed machine learning on Occlum
++  for (int i = 2; i < argc; ++i) {
++    char name[128], val[128];
++    if (sscanf(argv[i], "%[^=]=%s", name, val) == 2) {
++      //LOG(CONSOLE) << "setenv: " << std::string(name) << " = " << std::string(val);
++      setenv(std::string(name).c_str(), std::string(val).c_str(), 1);
++    }
++  }
++
+   rabit::Init(argc, argv);
+ 
+   std::vector<std::pair<std::string, std::string> > cfg;

demos/xgboost/run_xgboost.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+RED='\033[1;36m'
+NC='\033[0m'
+
+# Create a temporal folder and run xgboost demo
+tmp_dir="tmp_$RANDOM"
+mkdir -p $tmp_dir
+cp -a occlum_workspace/. $tmp_dir
+
+cd $tmp_dir
+echo -e "${RED}occlum run xgboost $@${NC}"
+occlum run /bin/xgboost /data/mushroom.conf $@

etc/template/Enclave.xml

@@ -4,7 +4,7 @@
   <ISVSVN>0</ISVSVN>
   <StackMaxSize>0x100000</StackMaxSize>
   <HeapMaxSize>0x2000000</HeapMaxSize>
-  <TCSNum>8</TCSNum>
+  <TCSNum>32</TCSNum>
   <TCSPolicy>1</TCSPolicy>
   <DisableDebug>0</DisableDebug>
   <MiscSelect>0</MiscSelect>