The adagrad branch contains code used for the SoCC'15 paper: Managed Communication and Consistency for Fast Data-Parallel Iterative Analytics
-
The system was tested on Ubuntu 14.04 and requires g++-4.8.
-
Install dependencies from Ubuntu's package manager:
sudo apt-get install uuid-dev, autoconf, libtool
sudo ln -s /usr/lib/x86_64-linux-gnu/librt.so /usr/lib/librt.so
- Download and build third-party dependencies:
make third_party_core
- Build the parameter server:
make ps_lib ml_lib
- Go to each application directory under apps to build applications
cd apps/matrixfact
make matrixfact_split
Before running the application, you need to prepare your dataset using data_split
which partitions the data file into N pieces (suppose you are using N machines) and convert it into a efficient binary format.
cd apps/lda
make lda