Monitoring HPC System Utilisation

Monitoring HPC System Utilisation

HPC and e-Science infrastructures are tasked with providing compute systems and services to, in many cases, a wide and varied range of scientific domains and/or research disciplines within domains. This often means that users of the systems and services have vastly varying levels of experience and that a wide and varied range of software applications are running on the systems.

To ensure that the provided systems are properly and well utilised, monitoring the utilisation of several types of resources (e.g. cpu, memory, etc.) can be essential. Examples of novice users allocating one or several multi-core compute nodes for a compute job, but only using one core are not uncommon and unless there is monitoring of the cpu utilisation on the system, then such jobs will not be easy to discover and follow up on. However, monitoring the degree of cpu utilisation for a given job allocation is only one example of why monitoring the utilisation of various types of resources can be useful. Other examples include finding challenging or poorly behaved jobs as well as learning what software applications that are running on the system and how much this software is used with respect to core-hours, number of users, cores per job, etc.

Knowledge mined from the collected data can clearly help guide support efforts and maybe even the design or procurement of new systems.

In this workshop, we will discuss what resources to monitor and what metrics to collect, how to collect the metrics, how to gather metrics in a cluster, how to store the data, and how to interface to the collected data. We will also discuss how to use the collected data to improve the utilisation of the system resources and identify problematic workloads, as well as what proactive support efforts this can lead to.

If you work with system operations and/or various types of user-oriented operational services aimed at helping users to utilise e-Science and HPC systems and services, then this workshop is for you.

Workshop format: Presentations followed by discussions

Workshop organizers:

  • Torben Rasmussen, NSC, SE
  • Bjørn-Helge Mevik, USIT, NO
  • Martin Dahlö, UPPMAX, SE


9:00 - 9:30 Presentation of our ideas with the workshop and potential items for the discussions


9:30 - 10:15 Thoughts and overview of various types of inefficiency in an academic HPC system

Peter Kjellström, NSC, SE

10:15 - 10:45 Coffee
10:45 - 11:15 Implementation and experiences from UPPMAX

Martin Dahlö, UPPMAX, SE

11:15 - 11:45 Proposal for software utilisation monitoring in SNIC


11:45 - 12:15 Software utilisation monitoring from NO

Henrik R. Nagel, NTNU, NO

12:30 - 13:30 Lunch
13:30 - 14:00 Experiences from NSC

Peter Bortas, NSC, SE

14:00 - 14:30 Presentation of Computerome dashboard

Erland Hochheim, DTU, DK

14:30 - 15:15 Monitoring of MPI lib usage at NSC

Peter Kjellström, NSC, SE

15:15 - 15:45 Coffee
15:45 - 16:15 Slurm-ganglia integration (Tromsø)

Roy Dragseth, UiT, NO

16:15 - 17:00 Discussions and closing conclusions


Workshop chair:
Torben Rasmussen

  • Monday, May 29, 09:00 - 12:30 (Room: Embla)
  • Monday, May 29, 13:30 - 17:00 (Room: Embla)