GPU-server “amp”

This manual is work in progress, please check regularly for updates

***Running your jobs on amp currently requires special procedure: you need to ssh into amp to submit jobs (and prepare your python envs***
The reason for this are differences in the operating system (amp: ubuntu, base: centos)



Hardware


amp

  • CPU: 2x AMD EPYC 7742 64core

  • RAM: 1 TB

  • GPUs: 8x A100 Nvidia 40GB

  • OS: Ubuntu

amp2

  • CPU: 2x AMD EPYC 7713 64core (3rd gen EPYC, Zen3)

  • RAM: 2 TB

  • GPUs: 8x A100 Nvidia 80GB

  • OS: Ubuntu





Login and $HOME


The server shares the home directory with all cluster nodes.

It is possible to login directly and submit and run jobs locally:

ssh uni-ID@amp.hpc.taltech.ee

Please don’t abuse the direct login by running jobs bypassing the queueing system. If this repeatedly disturbs jobs, we will have to disable direct login. Jobs for amp/amp2 need to be submitted from amp/amp2 (not from base) to ensure that all environment is set-up correctly. Anyway, jobs need to be submitted using srun or sbatch, do not run jobs outside the batch system.

Interactive jobs are started using srun:

srun -p gpu -t 1:00:00 --pty bash

GPUs have to be reserved/requested with:

srun -p gpu --gres=gpu:A100:1 -t 1:00:00 --pty bash

both amp/amp2 are in the same partition (-p gpu) so jobs that do not have specific requirements can run on any of the two nodes. If you need a specific type, e.g. for testing performance or because of memory requirements:

it is possible to request the feature “A100-40” (for the 40GB A100s in amp) or “A100-80” (for the 80GB A100s in amp2): --constraint=A100-40

another option is to request the job to run on a specific node, using the -w switch (e.g. srun -p gpu -w amp ... )

you can see which GPUs have been assigned to your job using echo $CUDA_VISIBLE_DEVICES, the CUDA-deviceID in your programs always start with “0” (no matter which physical GPU was assigned to you by SLURM).

The home-directory is the same as the cluster (/gpfs/mariana/home/<uniID>), additionally the AI-Lab home directories are mounted under /illukas/home/




Software and modules


modules specific on amp

enable the SPACK software modules (this is a separate SPACK installation from the rest of the cluster, use this one on amp):

module load amp-spack

the Nvidia HPC SDK (includes CUDA, nvcc and PGI compilers) is available with the following module directory (see below):

module load amp

from AI lab

enable the modules for AI lab software from illukas:

module load amp



GPU libraries and tools


The GPUs installed are Nvidia A100 with compute capability 80, compatible with CUDA 11. However, when developing own software, be aware of vendor lockin, CUDA is only available for Nvidia GPUs and does not work on AMD GPUs. Some new supercomputers (LUMI (CSC), El Capitan (LLNL), Frontier (ORNL)) are using AMD, and some plan the Intel “Ponte Vecchio” GPU (Aurora (ANL), SuperMUC-NG (LRZ)). To be future-proof, portable methods like OpenACC/OpenMP are recommended.

Porting to AMD/HIP for LUMI: https://www.lumi-supercomputer.eu/preparing-codes-for-lumi-converting-cuda-applications-to-hip/

Nvidia CUDA 11

Again, beware of the vendor lockin.

To compile CUDA code, use the Nvidia compiler wrapper:

nvcc

Offloading Compilers

  • PGI (Nvidia HPC-SDK) supports OpenACC and OpenMP offloading to Nvidia GPUs

  • GCC-10.3.0

  • GCC-11.2.0 with NVPTX supports GPU-offloading using OpenMP and OpenACC pragmas

  • LLVM-13.0.0 (Clang/Flang) with NVPTX supports GPU-offloading using OpenMP pragmas


See also: https://lumi-supercomputer.eu/offloading-code-with-compiler-directives/

OpenMP offloading

Since version 4.0 supports offloading to accelerators. It can be utilized by GCC, LLVM (C/Flang) and Nvidia HPC-SDK (former PGI compilers).

  • GCC-10.3.0

  • GCC-11.2.0 with NVPTX supports GPU-offloading using OpenMP and OpenACC pragmas

  • LLVM-13.0.0 (Clang/Flang) with NVPTX supports GPU-offloading using OpenMP pragmas

  • AOMP


List of compiler support for OpenMP: https://www.openmp.org/resources/openmp-compilers-tools/

Current recommendation: use Clang or GCC or AOMP

Nvidia HPC SDK

Compile option -⁠mp for CPU-OpenMP or -mp=gpu for GPU-OpenMP-offloading.

The table below summarizes useful compiler flags to compile you OpenMP code with offloading.

NVC/NVFortran Clang/Cray/AMD GCC/GFortran
OpenMP flag -mp -fopenmp -fopenmp -foffload=
Offload flag -mp=gpu -fopenmp-targets= -foffload=
Target NVIDIA default nvptx64-nvidia-cuda nvptx-none
Target AMD n/a amdgcn-amd-amdhsa amdgcn-amdhsa
GPU Architecture -gpu= -Xopenmp-target -march= -foffload=”-march=

OpenACC offloading

OpenACC is a portable compiler directive based approach to GPU computing. It can be utilized by GCC, (LLVM (C/Flang)) and Nvidia HPC-SDK (former PGI compilers).

Current recommendation: use HPC-SDK

Nvidia HPC SDK

Installed are versions 21.2, 21.5 and 21.9 (2021). These come with modulefiles, to use them, enable the the directory:

module load amp

then load the module you want to use, e.g.

module load nvhpc-nompi/21.5

The HPC SDK also comes with a profiler, to identify regions that would benefit most from GPU acceleration.

OpenACC is based on compiler pragmas enabling an incremental approach to parallelism (you never break the sequential program), it can be used for CPUs (multicore) and GPUs (tesla).

Compiling an OpenACC program with the Nvidia compiler: get accelerator information

pgaccelinfo

compile for multicore (C and Fortran commands)

pgcc -fast -ta=multicore -Minfo=accel -I/opt/nvidia/hpc_sdk/Linux_x86_64/21.5/cuda/11.3/targets/x86_64-linux/include/  -o laplace jacobi.c laplace2d.c
pgfortran -fast -ta=multicore  -Minfo=accel -I/opt/nvidia/hpc_sdk/Linux_x86_64/21.5/cuda/11.3/targets/x86_64-linux/include/ -o laplace_multicore laplace2d.f90 jacobi.f90

compile for GPU (C and Fortran commands)

pgcc -fast -ta=tesla -Minfo=accel  -I/opt/nvidia/hpc_sdk/Linux_x86_64/21.5/cuda/11.3/targets/x86_64-linux/include/ -o laplace_gpu jacobi.c laplace2d.c
pgfortran -fast -ta=tesla,managed -Minfo=accel -I/opt/nvidia/hpc_sdk/Linux_x86_64/21.5/cuda/11.3/targets/x86_64-linux/include/ -o laplace_gpu laplace2d.f90 jacobi.f90

Profiling:

nsys profile -t nvtx --stats=true --force-overwrite true -o laplace ./laplace
nsys profile -t openacc --stats=true --force-overwrite true -o laplace_data_clauses ./laplace_data_clauses 1024 1024

Analysing the profile using CLI:

nsys stat s laplace.qdrep

using the GUI:

nsys-ui

then load the .qdrep file.

GCC (needs testing)

  • GCC-10.3.0

  • GCC-11.2.0 with NVPTX supports GPU-offloading using OpenMP and OpenACC pragmas

HIP (upcoming)

For porting code to AMD-Instinct based LUMI, the AMD HIP SDK will be installed.




Example jobs


Example job for meshroom

** under construction **

Example job for colmap (needs GPU for part of the tasks)

** under construction **

See the AI-lab guide https://gitlab.cs.ttu.ee/ai-lab/samples/structure-from-motion-with-colmap, which needs to be slightly adapted.


Singularity

The container solution singularity is available (can also run docker container). Use with

module load amp
module load Singularity

pull the docker image you want, here ubuntu:18.04:

singularity pull docker://ubuntu:18.04

write an sbatch file (here called ubuntu.slurm):

#!/bin/bash
#SBATCH -t 0-00:30
#SBATCH -N 1
#SBATCH -c 1
#SBATCH -p gpu
#SBATCH --gres=gpu:A100:1
#SBATCH --mem-per-cpu=4000
singularity exec docker://ubuntu:18.04 cat /etc/issue

submit to the queueing system with

sbatch ubuntu.slurm

and when the resources become available, your job will be executed.