EvntConsistColl

Introduction

EvntConsistColl library is an extention to GASPI that provides some implementations of collectives, including eventually consistent ones, in addition to the GASPI native collectives like barrier and allreduce.

The main feature of the eventually consistent collectives is that they operate on a fraction of data (e.g. 60%) which makes them appealing for machine/ deep learning applications.

We provide implementations of broadcast, reduce, and allreduce.

Installation

Requirements:

• cmake version > 3.6 (presently build using cmake v3.9.3)
• c++ 14 (presently with gcc-8.2.0)

Building EvntConsistColl

1. clone the git-repository into <EvntConsistColl_root>

2. edit appropriatelly <EvntConsistColl_root>/CMakeFiles.txt to set the following variables

   • if GPI-2 is not to be loaded as a module, redefine PKG_CONFIG_PATH by adding the path to the file GPI2.pc (the package-config file for GPI-2)
   • eventually, comment the line with CMAKE_SHARED_LINKER_FLAGS ... (it has been added due to the relative old g++ system-libraries)

3. in <EvntConsistColl_root> create a subdirectory build to compile EvntConsistColl

   cd  <EvntConsistColl_root>
   mkdir build
   cd build
   cmake .. -DCMAKE_INSTALL_DIR=<target_installation_dir>
   make install
   

After building and installing EvntConsistColl,

• the library libEvntConsistColl.a is installed in <target_installation_dir>/lib
• the header EvntConsistColl.hxx is in <target_installation_dir>/include
• the executable examples are in <EvntConsistColl_root>/build/examples

Examples

There are few examples

• bcast provides two versions of broadcast with plain gaspi_write and binomial tree, while bcast_bench is primarily focused on benchmarking the later. Both support regular as well as eventually consistent collectives. To run bcast and bcast_bench (use binomial tree by default) inside build:

gaspi_run -m machine ./examples/bcast
gaspi_run -m machine ./examples/bcast_bench <number of elements> <iterations> [check, optional]

• reduce and reduce_bench provide implementations based on binomial tree that supports both regular and eventually consistent collectives. To run reduce and reduce_bench inside build:

gaspi_run -m machine ./examples/reduce <number of elements> <threshold in [0,1]>
gaspi_run -m machine ./examples/reduce_bench <number of elements> <iterations> [check, optional]

• allreduce_bench benchmarks the segmented pipelined ring implementation of allreduce. To run allreduce_bench inside build:

gaspi_run -m machine ./examples/allreduce_bench <number of elements> <iterations> [check, optional]