Skip to main content

PhD and habilitation thesis

My dissertation introduces a piece of software called Peano which is a certain way how to model tree-based adaptive Cartesian meshes and to arrange all data therein within stacks. That is: you can only run through this mesh in a depth-first manner along the Peano space-filling curve (SFC), but in return you use only data structures which are a fit to modern architectures as they are intrinsically cache optimal. Furthermore, you can encode a lot of adjacency and spatial information within the automaton which realises the traversal. So the memory footprint is relatively small. My software is not the first of its kind: there’s been a whole set of Peano codes at TUM at the time I started my research, but they all explored different avenues, and their core algorithm had been way more complicated than my approach: Indeed, they had to work with a number of stacks which is exponential in the spatial dimension d, whereas my algorithm is linear in d.

Part two and three of the thesis then use this tree traversal and stack paradigm to show how you can realise a multigrid algorithm on top of it which works totally matrix-free, and how you can parallelise this code based upon a multiscale generalisation of non-overlapping domain decomposition.

I call the resulting software Peano’s second generation. The first generation is the aforementioned set of codes. So this had been the first time, a lot of different Peano SFC ideas were condensated into one piece of code.

Nowadays, the book is available free of charge as

Obviously, you can still try to buy it from the publisher.

My habilitation extended the concepts of the PhD and also delivered a complete rewrite of Peano. I call this the third generation. it is also the generation you find in projects like ExaHyPE. The habilitation itself is cumulative, i.e. a collection of papers, but I wrote a rather long overview and contextualisation of these papers. While I cannot upload the papers themselves here (though most of them are gold access), you can download the habilitation’s intro here:

@tobiasweinzierl.bsky.social

  • Fifth day of Durham HPC Days. And still some people are up for a run in the early morning.
  • This is a picture of one of our #HPCDays morning runs. Was great, but the guys are (too) fit (for me).
  • Thanks @inseismoland.bsky.social for dropping by. An excellent conference depends on excellent talks! [contains quote post or other embedded content]
  • The first keynote is by Sven Bodo Scholz who starts from the observation that tuning makes code (quality) worse and unfit for modern, heterogeneous hardware. And therefore should be done by compilers where possible. But are our compilers and programming languages fit for purpose?
  • Totally full room for the first tutorial of the HPC Day run by friends of Nvidia on AI model upscaling.