Different workloads on HPC

Europe/Warsaw
online

online

Klemens Noga (ACC Cyfronet AGH), Oskar Klimas (ACK Cyfronet AGH)
Description

The training focuses on different ways to allocate resources on supercomputers and parallelisation schemes used to speed up calculations

Modern supercomputers are massive parallel machines with up to hundreds of cores on node and even thousands of nodes connected via fast network interface. To utilise fully those machines, one needs to know about parallelisation techniques and queuing system options of resource allocation. This training will show different ways that SLURM allocates resources and how to apply them to different computations and workloads.

Agenda

  • Serial jobs
  • Embarrassingly parallel jobs
  • Intranode calculations
    • Types of parallelisation 
    • Granularity of resource allocation 
  • Multinode calculations 
  • Gpu calculations
  • Array jobs - high throughput jobs
  • Job dependencies
  • Brief review of available applications that use described methods of parallelisation 

Duration

3 hours

Target audience

Training is intended for users who have started running their own calculations and need to learn about different methods of allocating resources on HPC clusters. 

Requirements

We recommend doing this training after completing Running Calculations the HPC way training provided at ACK Cyfronet AGH.

Technical prerequisites:

The above prerequisites are required to actively participate in the hands-on exercises.
In case of no active affiliation or PLGrid account, please contact the organisers at training@cyfronet.pl

Participants' computers should be equipped with software that allows SSH connections. On Windows, it is recommended to use Windows Terminal after checking if OpenSSH is installed. We advise against using Putty.

Venue

The workshop will be conducted online on the Zoom platform. The meeting link will be sent to registered participants.

Language

English/Polish - dependent on the participants.

Level

Basic

Organised by

ACC Cyfronet AGH

    • 09:00 09:15
      Introduction 15m

      Introduction to SLURM job scheduler.

    • 09:15 09:35
      Serial jobs and embarrassing parallel calculations 20m

      How to perform serial computations and how to run concurrently multiple separate calculations within one SLURM job.

    • 09:35 10:20
      Intranode calculations 45m

      How to perform calculations within a single node with different parallelisation frameworks, and how to use SLURM to allocate resources

    • 10:20 10:30
      coffee break 10m
    • 10:30 11:15
      Multinode calculations 45m

      How to perform calculations that use resources of more than one node. Especially comuptations which use MPI or MPI+OpenMP frameworks.

    • 11:15 11:45
      Gpu calculations 30m

      How to run computations which use GPUs.

    • 11:45 12:00
      Arrays of jobs and dependencies between jobs 15m

      How to schedule many jobs with single SLURM submit script and how to make dependenices between jobs