HPL.

HPL is a software package that solves a (random) dense linear system in double precision (64 bits) arithmetic on distributed-memory computers. It can thus be regarded as a portable as well as freely available implementation of the High Performance Computing Linpack Benchmark.

The algorithm used by HPL can be summarized by the following keywords: Two-dimensional block-cyclic data distribution – Right-looking variant of the LU factorization with row partial pivoting featuring multiple look-ahead depths – Recursive panel factorization with pivot search and column broadcast combined – Various virtual panel broadcast topologies – bandwidth reducing swap-broadcast algorithm – backward substitution with look-ahead of depth 1.

Official website for HPL : https://www.netlib.org/benchmark/hpl/

Building HPL using Ampere Oracle Blis Libraries.

Building HPL with Ampere-Oracle Blis libraries is very easy and should not take a lot of time. It’s a 2-step process: Step 1: Where we build the Math libraries found on the Ampere branch of Oracle Blis libraries and Step 2 where we build the HPL binaries.

A detailed guide is below:

Step 1: Prerequisites:

System Config :

OS : Ubuntu 20.04

GCC : 12.2.0

Kernel : 5.4.0-148-generic

To ensure a seamless build process, both, the math libraries and the benchmark are built inside the /opt directory.

a. Downloading and installing Ampere Oracle Blis Libraries:

pushd /opt
git clone https://github.com/flame/blis.git MyBlisDir
pushd MyBlisDir
#Switch to the new ampere branch 
git checkout ampere
./QuickStart.sh altramax

Ensure that the test bench contains Ampere Oracle Blis exported to PATH and LD_LIBRARY_PATH appropriately.

source ./blis_build_altramax.sh
source blis_setenv.sh
export LD_LIBRARY_PATH=/opt/MyBlisDir/lib/altramax
popd
popd

b. OpenMPI: Along with Ampere Oracle Blis, we will also need openmpi. We have used openmpi 4.1.4. An installation guide for openmpi can be found inside the tarball: https://download.open-mpi.org/release/open-mpi/v4.1/openmpi-4.1.4.tar.gz

If OpenMPI is installed in a non-default location. Add the directory location to PATH and the directory location to LD_LIBRARY_PATH using the following commands

export PATH=<PATH_TO_OPENMPI_BIN_DIR>:$PATH
export LD_LIBRARY_PATH=< PATH_TO_OPENMPI_LIB_DIR>:$LD_LIBRARY_PATH

Ensure successful installation of openmpi by executing the following commands.

mpirun --version  #(That should bring up the openmpi version. 4.1.4 in this case)
mpicc --version #(That should bring up the installed gcc version)
mpic++ --version #(That should bring up the installed g++ version)
mpifort --version #(That should bring up the installed gfortran version)

If any of the above 3 commands do not return the version for gcc/g++/gfortran, install the missing gcc/g++/gfortran for your distro using

“sudo apt/yum/dnf install <package_name>”

Step 2: Building HPL Benchmark

a. Downloading and Installing HPL 2.3

pushd /opt
wget https://netlib.org/benchmark/hpl/hpl-2.3.tar.gz
tar -xzf hpl-2.3.tar.gz
popd

Copy the Makefile attached with this document to /opt/hpl-2.3 folder.

cp Make.Altramax_oracleblis /opt/hpl-2.3

Compile the HPL binary

make arch=Altramax_oracleblis -j

Upon success, a bin folder will be created. This folder should contain 2 files: xhpl (which is the HPL #binary) and HPL.dat (which is the standard input file).

pushd /opt/hpl-2.3/bin/Altramax_oracleblis

Step 3: Creating the HPL input file

Sample HPL.dat file attached.

Copy the attached HPL.dat file to “/opt/hpl-2.3/bin/Altramax_oracleblis “

Please note that this HPL.dat file is designed to run on 96 cores at 64 GB RAM. If you have access to more RAM please refer to Step 5 on how to maximise the values for Ns

Step 4: Run the benchmark

mpirun -np 96 --bind-to core --map-by core ./xhpl &> out.log

Step 5: Performance Expectations

If your system differs from our testbench, the HPL.dat file will need to be modified (line #6) to match your respective Altra Max config.

Table 1: Showcasing the line which needs to be modified in HPL.dat

Line Number	Value	Description
6	150000	Ns

The value of N when changed to 150K should take approximately 180 gigs of memory and would run on a machine having 256 GB memory. Table2 with differing values of Ns is shown below with our reference numbers.

Table 2: Varying sizes on Ns w.r.t available system memory.

Our observed Results for AltraMax 96 cores @ 2.8GHz:

Input Param	Input Param
NB=256	P=8 Q=12

Problem Size (Ns)	Memory Used	Results (Gflops)
70K	46 gig	1151
100K	83 gig	1250
105K	91 gig	1253

Our observed Results for AltraMax 128 cores @ 3.0GHz:

Input Param	Input Param
NB=256	P=8 Q=16

Problem Size (Ns)	Memory Used	Results (Gflops)
150K	177 gig	1528
200k	312 gig	1552
250k	480 gig	1597

joespeed / HPL-ADLINK