"Finding the best environment for containers"

High-performance computing will soon be unthinkable without containers. The digital boxes contain all the tools needed to run a scientific code, but they do not perform equally well on all computer systems. The reasons for this are being researched and evaluated at the Leibniz Computing Centre.

 

Running scientific code on multiple supercomputers efficiently is a challenge. High-performance and supercomputers are custom-built, differ from their counterparts in many technical details, and therefore do not fit every programme. Containers, however, can build bridges between systems: They are designed to improve the portability of applications on supercomputers. To do this, hardware such as processors is virtualised on container platforms, and the digital boxes also contain all the tools needed to run code, such as the necessary kernels, software libraries and the application itself. However, the container strategy has recently begun to show performance limitations: "The number of heterogeneous computer architectures is growing, and given the variety of requirements and applications, we are sometimes faced with the question of which system is best suited to run an application," says Maximilian Höb, a computer scientist at the Leibniz Computing Centre (LRZ). "Containers cannot therefore deliver the same performance on every high-performance computer". Höb is investigating these differences in his doctoral thesis, researching container technologies and developing a model to evaluate their performance data from different perspectives. His work has just been acknowledged at the Supercomputing Asia 2024 (SCA24) in Sydney.

Congratulations on winning the Doctoral Thesis Award at Supercomputing Asia: What did you present there? Maximilian Höb: During one session I gave a pitch for my PhD thesis. In a three-minute presentation, I tried to explain what my work is about. It is about analysing performance data for containerised applications in HPC.

What does the prize mean to you? Höb: First of all, of course, it is a great honour. It confirms the importance of the topic of my work and and it also recognises my effort - I have succeeded in finding a common thread for this broad topic and in dealing with it in a concise way.

Why are you interested in containerisation in the data centre? Höb: A project at the Ludwig Maximilian University in Munich introduced me to the world of containers. The more I looked into it, the more I realised that users will use containers for their own code, especially in the area of high performance computing. The last few years have shown that containerisation has indeed become a trend, that there are more and more technologies for it, and that interest in it is growing rapidly.

How do you research for your thesis and what is your focus? Höb: In any PhD thesis, theory is of course the main focus. That's why I'm working on a model and a method to evaluate it. This means that I am determining and comparing different performance characteristics. This, in turn, helps to run containers on the system for which they are best suited. The number of heterogeneous computer architectures is growing and, and given the variety of requirements and applications, we are sometimes faced with the question of which system is best suited to run an application, but also whether the resources provided are actually needed or whether their use can be optimised. Containers, for example, may not deliver the same performance on every high performance machine.

Can containers also help reduce the energy requirements of a data centre or supercomputer? How? Höb: We should create an optimal environment for every container. The term "optimal" can be defined in many ways - for example, as a short runtime or as an efficient, energy-saving design. Those who operate a system and define the objectives determine what is optimal. In my approach, important parameters are modelled in a comparable way, which then offers many possibilities for optimisation.

What happens next with the dissertation? Höb: My aim is to submit the thesis in the next few months. To do this, I need to revise the model and its evaluation, as well as evaluate the parameters in a broader sense. The aim of my thesis is not to develop a productive system, but rather to build the theoretical model for it and confirm it with a proof of concept. (Interview: vs, ssc)

M. Höb (2nd from the left) at the award ceremony in Sydney with Prof. Dieter Kranzlmüller (left),
head of LRZ, and Prof. Ute Roessner and Prof. Sean Smith, both from
Australian National University