22 days ago
High Performance Analytics Toolkit (HPAT) is a Julia-based framework for big data analytics on clusters that is both easy to use and extremely fast; it is orders of magnitude faster than alternatives like Apache Spark*.
HPAT automatically parallelizes analytics tasks written in Julia, generates efficient MPI/C++ code, and uses existing high performance libraries such as HDF5 and Intel® Data Analytics Acceleration Library (Intel® DAAL). HPAT is based on ParallelAccelerator and CompilerTools packages.
HPAT is in early development and therefore feedback is highly appreciated.
Presentation at JuliaCon 2016:
$ julia -e 'Pkg.add("HPAT")' $ mpirun -np 2 julia ~/.julia/v0.4/HPAT/examples/pi.jl --points=1000000
On Ubuntu, these commands resolve some MPI.jl issues:
$ sudo apt-get install -y g++ gcc gfortran cmake openmpi-bin openmpi-common libopenmpi-dev libhdf5-openmpi-dev $ julia ~/.julia/v0.4/MPI/deps/build.jl
Front page of Spark* website demonstrates over two orders of magnitude speedup compared to Hadoop* on Logistic Regression example. Simply put, Spark* keeps data in memory while Hadoop* reads and writes to disks frequently.
HPAT is two orders of magnitude faster than Spark*! Data is kept in processor registers as much as possible with HPAT, which is necessary for best performance. In addition, HPAT doesn't have Spark*'s TCP/IP and Java Virtual Machine (JVM) overheads since it generates "bare-metal" MPI/C++ code.
Here is how one can compare the performance of HPAT and Spark* for Logistic Regression example on a local Ubuntu machine. Note that parallel HDF5 is required.
Install Julia and dependencies:
$ sudo add-apt-repository ppa:staticfloat/juliareleases $ sudo add-apt-repository ppa:staticfloat/julia-deps $ sudo apt-get update $ sudo apt-get install -y gcc g++ gfortran cmake openmpi-bin openmpi-common libopenmpi-dev libhdf5-openmpi-dev julia libblas-dev libopenblas-dev $ julia -e 'Pkg.add("HPAT")'
Generate input data:
# generate data with 100 million labeled instances $ julia -e 'Pkg.add("HDF5")' $ julia $HOME/.julia/v0.4/HPAT/generate_data/generate_logistic_regression.jl --instances=100000000 --path=/tmp/
Run Logistic Regression example of HPAT:
# run on 64 MPI processes $ mpirun -np 8 julia $HOME/.julia/v0.4/HPAT/examples/logistic_regression.jl --iterations=200 --file=/tmp/logistic_regression.hdf5 &> lr_hpat.out
Run Logistic Regression example of Spark*:
# assuming spark is configured properly (with driver and executor memory set) $ spark-submit $SPARK_HOME/examples/src/main/python/logistic_regression.py /tmp/logistic_regression.csv 200 &> lr_spark.out
Monte Carlo Pi estimation is another interesting example of Spark*. HPAT is over 1000 times faster for this case for various reasons! First, HPAT can divide computation statically since it generates code rather than executing the program operation-by-operation. Spark* uses a dynamic scheduler with high overhead. But more importantly, Spark* generates an enormous array for the map operation, then executes the reduce operation. However, HPAT uses ParallelAccelerator which removes unnecessary arrays. Therefore, HPAT doesn't create any arrays and the generated code is just a simple loop. Hence, the computation is performed in registers and there is no extra memory access.
Run Pi example of HPAT:
$ mpirun -np 64 julia ~/.julia/v0.4/HPAT/examples/pi.jl --points=10000000
Run Pi example of Spark*:
$ spark-submit $SPARK_HOME/examples/src/main/python/pi.py 100 &> pi_spark.out
Programmers need to follow these directions in order to use HPAT effectively:
Currently, there is a high initialization cost (e.g. 25s) the first time HPAT is used within a Julia runtime since package precompilation of Julia v0.4 is very limited. We are working on various solutions.