Overview

Katana is a library and set of applications to work with large graphs efficiently.

Highlights include:

• Parallel graph analytics and machine learning algorithms
• Multithreaded execution with load balancing and excellent scalability on multi-socket systems
• Interfaces to build your own algorithms
• Scalable concurrent containers such as bag, vector, list, etc.

Katana is released under the BSD-3-Clause license.

Installing

The easiest way to get Katana is to install the Conda packages. See the Conda User Guide for details on how to install Conda.

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

Then, install Katana

conda install -c katanagraph/label/dev katana katana-python

Currently, only development versions are available through Conda. These builds reflect the current state of the master branch. Releases of stable versions will be available in the future.

Alternatively, you can build Katana from source.

Building from Source

See the build instructions for details.

Dependencies

Katana is tested on 64-bit Linux, specifically Ubuntu LTS 20.04.

At the minimum, Katana depends on the following software:

• A modern C++ compiler compliant with the C++-17 standard (gcc >= 7, Intel >= 19.0.1, clang >= 7.0)
• CMake (>= 3.17)
• Boost library (>= 1.58.0, we recommend building/installing the full library)
• libllvm (>= 7.0 with RTTI support)
• libfmt (>= 4.0)
• libxml2 (>= 2.9.1)
• libarrow (>= 4.0)
• libuuid (>= 2.31.1)
• nlohmann/json (>= 3.7.3)

Here are the dependencies for the optional features:

• Linux HUGE_PAGES support (please see www.kernel.org/doc/Documentation/vm/hugetlbpage.txt). Performance will most likely degrade without HUGE_PAGES enabled. Katana uses 2MB huge page size and relies on the kernel configuration to set aside a large amount of 2MB pages. For example, our performance testing machine (4x14 cores, 192GB RAM) is configured to support up to 65536 2MB pages:

  cat /proc/meminfo | fgrep Huge
  AnonHugePages:    104448 kB
  HugePages_Total:   65536
  HugePages_Free:    65536
  HugePages_Rsvd:        0
  HugePages_Surp:        0
  Hugepagesize:       2048 kB

• libnuma support. Performance may degrade without it. Please install libnuma-dev on Debian like systems, and numactl-dev on Red Hat like systems.

• Doxygen (>= 1.8.5) for compiling documentation as webpages or latex files

• PAPI (>= 5.2.0.0 ) for profiling sections of code

• Vtune (>= 2017 ) for profiling sections of code

• Eigen (3.3.1 works for us) for some matrix-completion app variants

Documentation

To build documentation for this project, see the documentation instructions.

Using Katana as a Library

There are two common ways to use Katana as a library. One way is to copy this repository into your own CMake project, typically using a git submodule. Then you can put the following in your CMakeLists.txt:

add_subdirectory(galois EXCLUDE_FROM_ALL)
add_executable(app ...)
target_link_libraries(app Katana::galois)

The other common method is to install Katana outside your project and import it as a package.

If you want to install Katana, assuming that you wish to install it under INSTALL_DIR:

cmake -DCMAKE_INSTALL_PREFIX=$INSTALL_DIR $SRC_DIR
make install

Then, you can put something like the following in CMakeLists.txt:

list(APPEND CMAKE_PREFIX_PATH ${INSTALL_DIR})
find_package(Katana REQUIRED)
add_executable(app ...)
target_link_libraries(app Katana::galois)

If you are not using CMake, the corresponding basic commands (although the specific commands vary by system) are:

c++ -std=c++14 app.cpp -I$INSTALL_DIR/include -L$INSTALL_DIR/lib -lkatana_galois

Contact Us

For bugs, please raise an issue on GiHub.

If you file an issue, it would help us if you sent (1) the command line and program inputs and outputs and (2) a core dump, preferably from an executable built with the debug build.

You can enable core dumps by setting ulimit -c unlimited before running your program. The location where the core dumps will be stored can be determined with cat /proc/sys/kernel/core_pattern.

To create a debug build, assuming you will build Katana in BUILD_DIR and the source is in SRC_DIR:

cmake -S $SRC_DIR -B $BUILD_DIR -DCMAKE_BUILD_TYPE=Debug
make -C $BUILD_DIR

A simple way to capture relevant debugging details is to use the script command, which will record your terminal input and output. For example,

script debug-log.txt
ulimit -c unlimited
cat /proc/sys/kernel/core_pattern
make -C $BUILD_DIR <my-app> VERBOSE=1
my-app with-failing-input
exit

This will generate a file debug-log.txt, which you can supply when opening a GitHub issue.