xgboost is compiled in local mode

[antonymayi]

Hi,

really struggling to compile xgboost for yarn with hdfs support. I can launch the yarn app but in the container logs it show:

terminate called after throwing an instance of 'std::runtime_error'
  what():  xgboost is compiled in local mode
to use hdfs, s3 or distributed version, compile with make dmlc=1

The libhdfs library on my system is placed as follows:

/usr/lib64/libhdfs.so
/usr/lib64/libhdfs.so.0.0.0
/usr/lib/hadoop/lib/native/libhdfs.a
/usr/include/hdfs.h

I tried building xgboost using these options:

make USE_HDFS=1 HDFS_LIB_PATH=/usr/lib64 JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64 HADOOP_HDFS_HOME=/usr
make USE_HDFS=1 HDFS_LIB_PATH=/usr/lib/hadoop/lib/native/ JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64 HADOOP_HDFS_HOME=/usr
make USE_HDFS=1 HDFS_LIB_PATH=/usr/lib64 HDFS_INC_PATH=/usr/include JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64
make USE_HDFS=1 HDFS_LIB_PATH=/usr/lib/hadoop/lib/native/ HDFS_INC_PATH=/usr/include JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64

But still no joy, keep getting the "xgboost is compiled in local mode".

Can you please advise?

Thanks,
Antony.

[antonymayi]

ok, I managed to build it using:

make USE_HDFS=1 HDFS_LIB_PATH=/usr/lib64 HDFS_INC_PATH=/usr/include JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64 dmlc=../dmlc-core/

however that makes it dependent on libglog so I had to hack repo/xgboost/Makefile:

 @@ -2,7 +2,7 @@ export CC  = $(if $(shell which gcc-5),gcc-5,gcc)
 export CXX = $(if $(shell which g++-5),g++-5,g++)
 export MPICXX = mpicxx
-export LDFLAGS= -pthread -lm
+export LDFLAGS= -pthread -lm -lglog -L../../deps/lib

This however now makes the yarn app dependent on lobglog.so I have to submit dmlc_yarn.py with -u deps/lib/libglog.so.0 which is bit dirty.

[antonymayi]

so final method to get it all working without the libglog dependency:

cd repo/xgboost
git clone https://github.com/dmlc/dmlc-core
cd dmlc-core
make USE_HDFS=1 HDFS_LIB_PATH=/usr/lib64 HDFS_INC_PATH=/usr/include JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64
cd ..
make USE_HDFS=1 HDFS_LIB_PATH=/usr/lib64 HDFS_INC_PATH=/usr/include JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64 dmlc=dmlc-core
cp xgboost ../../bin/xgboost.dmlc

this is rather a hack/workaround, it is not straightforward build process given the declared wormhole build config options.

[Kornel]

I'm having a similar issue, the tricks to tweak the build helped, however now I'm getting error like these:

Container: container_1447066627326_186565_01_000004 on s40819.dc4.local_8041
==============================================================================
LogType: stderr
LogLength: 130
Log Contents:
./xgboost.dmlc: error while loading shared libraries: libhdfs.so.0.0.0: cannot open shared object file: No such file or directory

LogType: stdout
LogLength: 0
Log Contents:

Well, the quick fix is to add libhdfs.so.0.0.0 using the -u option in dmlc_yarn.py .

Now my logs are clean

Container: container_1447066627326_186575_01_000028 on s40834.dc4.local_8041
==============================================================================
LogType: stderr
LogLength: 0
Log Contents:

LogType: stdout
LogLength: 0
Log Contents:

but the application still fails! I've attached the logs I've got from yarn logs -applicationId

Now I'm lost, hard to guess where to go from here. My guess is, something is wrong that I had to upload the libhdfs.so.0.0.0 file?

Any ideas?

logs.txt