Mike Merrill's CHIUW 2019 talk
Exploratory data analysis (EDA) is a prerequisite for all data science, as illustrated by the ubiquity of Jupyter notebooks, the preferred interface for EDA among data scientists. The operations involved in exploring and transforming the data are often at least as computationally intensive as downstream applications (e.g. machine learning algorithms), and as datasets grow, so does the need for HPC-enabled EDA. However, the inherently interactive and open-ended nature of EDA does not mesh well with current HPC usage models. Meanwhile, several existing projects from outside the traditional HPC space attempt to combine interactivity and distributed computation using programming paradigms and tools from cloud computing, but none of these projects have come close to meeting our needs for high-performance EDA.
To fill this gap, we have developed a software package, called Arkouda, which allows a user to interactively issue massively parallel computations on distributed data using functions and syntax that mimic NumPy, the underlying computational library used in the vast majority of Python data science workflows. The computational heart of Arkouda is a Chapel interpreter that accepts a pre-defined set of commands from a client (currently implemented in Python) and uses Chapel's built-in machinery for multi-locale and multithreaded execution. Arkouda has benefited greatly from Chapel's distinctive features and has also helped guide the development of the language.
In early applications, users of Arkouda have tended to iterate rapidly between multi-node execution with Arkouda and single-node analysis in Python, relying on Arkouda to filter a large dataset down to a smaller collection suitable for analysis in Python, and then feeding the results back into Arkouda computations on the full dataset. This paradigm has already proved very fruitful for EDA. Our goal is to enable users to progress seamlessly from EDA to specialized algorithms by making Arkouda an integration point for HPC implementations of expensive kernels like FFTs, sparse linear algebra, and graph traversal. With Arkouda serving the role of a shell, a data scientist could explore, prepare, and call optimized HPC libraries on massive datasets, all within the same interactive session.
- requires chapel 1.20.0 with the --legacy-classes flag
- requires zeromq version >= 4.2.5, tested with 4.2.5 and 4.3.1
- requires python 3.6 or greater
- requires numpy
- requires Sphinx and sphinx-argparse to build python documentation
brew install chapel
# you can also install python3 with brew
brew install python3
# the arkouda python client is available via pip
# pip will automatically install python dependencies (zmq and numpy)
# however, pip will not build the arkouda server (see below)
pip3 install arkouda
# these packages are nice but not a requirement
pip3 install pandas
pip3 install jupyter# on my mac build chapel in my home directory with these settings...
export CHPL_HOME=~/chapel/chapel-1.20.0
source $CHPL_HOME/util/setchplenv.bash
export CHPL_COMM=gasnet
export CHPL_COMM_SUBSTRATE=smp
export CHPL_TARGET_CPU=native
export GASNET_QUIET=Y
export CHPL_RT_OVERSUBSCRIBED=yes
cd $CHPL_HOME
makeDownload, clone, or fork the arkouda repo. Further instructions assume that the current directory is the top-level directory of the repo.
If your environment requires non-system paths to find dependencies (e.g.,
if using the ZMQ and HDF5 bundled with Anaconda), append each path to a new file Makefile.paths like so:
# Makefile.paths
# Custom Anaconda environment for Arkouda
$(eval $(call add-path,/home/user/anaconda3/envs/arkouda))
# ^ Note: No space after comma.The chpl compiler will be executed with -I, -L and an -rpath to each
path.
Now, simply run make to build the arkouda_server executable.
Make sure you installed the Sphinx and sphinx-argparse packages (e.g. pip3 install -U Sphinx sphinx-argparse)
Run make doc to build both the Arkouda python documentation and the Chapel server documentation
The output is currently in subdirectories of the arkouda/doc
arkouda/doc/python # python frontend documentation
arkouda/doc/server # chapel backend server documentation
To view the documentation for the Arkouda python client, point your browser to file:///path/to/arkouda/doc/python/index.html, substituting the appropriate path for your configuration.
The command-line invocation depends on whether you built a single-locale version (with CHPL_COMM=none) or multi-locale version (with CHPL_COMM set).
Single-locale startup:
./arkouda_serverMulti-locale startup (user selects the number of locales):
./arkouda_server -nl 1By default, the server listens on port 5555 and prints verbose output. These options can be changed with command-line flags --ServerPort=1234 and --v=false.
There is a small test program that connects to a running arkouda_server, runs a few computations, and shuts down the server. To run it, open a new terminal window in the arkouda directory and run
python3 tests/check.py localhost 5555Substitute the correct hostname and port if you used a different configuration.
If you'd like to contribute, please see CONTRIBUTING.md.