commKernel

MPI left-right communication kernel for performance testing on Stampede. Three binaries are used for testing:

threadWorkComm

Only 1 MPI task is ran and the number of workers corresponds to the number of pthreads that are created. Data is exchanged between pthreads. This code is using MPI_THREAD_MULTIPLE

mpiWork

Here workers are representing MPI tasks. Data is exhchanged between MPI tasks. The code is using MPI_INIT.

mpiWorkThreadMult

Here workers are representing MPI tasks. Data is exhchanged between MPI tasks. The code is using MPI_THREAD_MULTIPLE.

tl;dr

MPI_THREAD_MULTIPLE peformance with pthreads is bad on the Phi.

The following presentation and paper discuss the use of a library that implements the proposed MPI endpoints apis to increase bandwidth at the cost of increased latency. Results show net application speedup despite the latency penalty.

http://meetings.mpi-forum.org/secretary/2014/12/slides/Hybrid-Plenary--EP-Lib.pdf

http://pcl.intel-research.net/publications/sridharan-sc14.pdf

contents

• envImpi.sh - environment file for Intel MPI 4
• envImpi5.sh - environment file for Intel MPI 5
• envImpi51.sh - environment file for Intel 16 and Intel MPI 5.1 (beta)
• envMvapich.sh - environment file for Mvapich2-mic
• envMvapich2.sh - environment file for Mvapich2-mic v2.0 - edit the install path
• envMpich3.sh - environment file for MPICH 3.2b2 - edit the install path
• build.sh - build script
• build.x86.sh - build script for stampede host processors
• getTimeImpi.sh - run script for Intel MPI 4 and 5
• getTimeImpi.x86.sh - run script for Intel MPI 4 and 5 on stampede host processors
• getTimeImpi51.sh - run script for Intel MPI 5.1
• getTimeMvapich.sh - run script Mvapich2-mic
• getTimeMvapich2.sh - run script Mvapich2-mic
• getTimeMpich3.sh - run script MPICH 3.2b2
• kernelComm.c - communication kernel
• kernelComm.x86.c - communication kernel with affinity on stampede host processors
• kernel.h - kernel header
• mpiWork.c - mpi driver
• README.md - this file
• threadWork.c - threaded mpi driver

build

source env<Impi|Impi51|Mvapich|Mvapich2|Mpich3>.sh
for i in 1 2 4; do ./build[.x86].sh $i; done

run

From an interactive idev session:

./getTime<Impi|Mvapich|Mvapich2|Mpich3>[.x86].sh

Performance Results

The following tests used MPSS 3.3 as installed on Stampede. see performance.csv

Mvapich2-MIC v2.0

Mvapich2-MIC v2.0 is not installed on Stampede. To install it follow the instructions below:

download

wget http://mvapich.cse.ohio-state.edu/download/mvapich/mic/mvapich2-mic-2.0_mpss-3.3.run

run installer

./mvapich2-mic-2.0_mpss-3.3.run
#set the install path to something in your home/work directory
#accept defaults for library paths 

edit the 'ins' variable in the environment file 'envMvapichMic2.sh' to point at the chosen install path

sanity check the install by running the latency benchmark

source envMvapich2.sh

create file 'config'

echo "-n 2 : $MV2_MIC_INSTALL_PATH/libexec/mvapich2/osu_latency" > config

create file 'hosts'

echo 'mic0:2' >> hosts

run

mpirun_rsh -config ./config -hostfile ./hosts

MPICH3

MPICH 3.2b is not installed on Stampede. To install it follow the instructions below:

download

wget http://www.mpich.org/static/downloads/3.2b2/mpich-3.2b2.tar.gz

unpack, configure, and build

tar xzf mpich-3.2b2.tar.gz
cd mpich-3.2b2
mkdir buildIntel15
cd buildIntel15
./doConfigIntel15.sh <install prefix absolute path>
make 
make install