Skip to content

Latest commit

 

History

History
156 lines (110 loc) · 3.7 KB

2020-12-19meeting.md

File metadata and controls

156 lines (110 loc) · 3.7 KB
Raw

What is ASC Student Supercomputer Challenge

exp4.jpg

Application Performance Optimization

  1. Compile & Run
  • Read the installation documentation
  • Compile and install
  • Find optimal parameters for running
  • Compare the performance of difference of libraries and find the best library
  1. Analysis of computing resource allocation efficiency

​ Number of cores -> Number of cores per node -> Number of nodes

  1. Application performance analysis

  • Intel Compiler analysis

    • -profile-functions
    • -profile-loops=all

  • Use Intel VTune to detect hotspots

    Intel Vtune

    Use it to get a performance snapshot

    • MPI load balancing
    • OpenMP load balancing
    • Memory stalls
    • FPU Utilization
    • I/O

  • Turn off IPO and function inline while analyzing performance

  1. Compilation Optimization

Build without optimization -> General optimization(-o1 -o2 -o3) -> Architecture related optimization -> Interprocedural Optimization -> Vectorization -> Automatic Parallelization

for (i=0; i<n; i++)
    c[i] = a[i] + b[i];
    
c1=a1+b1
c2=a2+b2
...
  1. Advanced Optimization
  • Read algorithm documentation
  • Read the source code and find parallelizable areas
  • Parallelize the hotspots areas according to performance analysis results
    • OpenMP
    • OpenACC
    • CUDA
    • Intel Cilk
    • TBB
  • Balance the workload of MPI processes and threads

HPCG as An Example

Installation Environment:

Ubuntu 18.04.2 LTS, 4.18.0-15-generic, x86_64, VirtualBox, 2GB Memory, HPCG 3.1

0. Install dependency

sudo apt-get install -y mpich

1. Download HPCG

cd ~
wget http://www.hpcg-benchmark.org/downloads/hpcg-3.1.tar.gz
gunzip hpcg-3.1.tar.gz; tar -xvf hpcg.tar

2. Generate HPCG template conffile

# Working Directory: hpcg-3.1/setup
cd hpcg-3.1/setup
cp Make.Linux_MPI Make.linux
# These Files will be Generated:
	# hpcg-3.1/setup/Make.linux		# Compilation Configuration File

3. Modify Make.linux according to the text below

MPdir        = /usr/bin
MPinc        = -I /usr/include/mpich/
MPlib        = /usr/lib/x86_64-linux-gnu/libmpich.so

4. Compile HPCG

# Working Directory: hpcg-3.1/
mkdir build
cd build
../configure linux
make -j $(nproc)
# These Files will be Generated:
	# hpcg-3.1/build/bin/hpcg.dat		# Benchmark Configuration File
	# hpcg-3.1/build/bin/xhpcg			# Executable File

5. Modify hpcg.dat and run the HPCG

Modify hpcg.dat based on the text below (It dependes on your memory size):
HPCG benchmark input file
Sandia National Laboratories; University of Tennessee, Knoxville
16 16 16
60
  • 16 16 16 is problem size, limited by system Memory.
  • 60 is test time. 60 for optimization, 1860 for official performance test.
Execute HPCG program
# Working Dir: hpcg-3.1/build/bin
mpirun -np 4 ./xhpcg
# These Files will be Generated:
	# hpcg-3.1/build/bin/hpcg20191202T131917.txt
	# hpcg-3.1/build/bin/HPCG-Benchmark_3.1_2019-12-02_13-20-15.txt
Output Explain
Final Summary::HPCG result is VALID with a GFLOP/s rating of=13.9936

Package Managing

Using Spack!

A flexible package manager that supports multiple versions, configurations, platforms, and compilers.

spack install mpileaks
spack install [email protected] %[email protected] +threads cppflags="-O3 -g3" target="skylake" ^mpich3.3 // customer. version, compiler, build options, compiler flags, target mircoarchitecture, dependencies

Write python code and yaml to build your own package.

Planning

  1. Get familiar with working environment
  2. Learn basic MPI, OpenMP
  3. Run benchmarks and choose one to optimize it.