steffenseckler/tum-dissertation-latex repository overview

⁠Dissertation

Newest pdf can be downloaded here⁠

⁠Gliederung

Algorithm and Performance Engineering for HPC Particle Simulations

⁠Abstract

• most optimizations on ls1-mardyn, but applicable elsewhere.

⁠Introduction

1. Motivation, warum ist das, was ich mache wichtig.
2. Particle Simulations -- short range
   • MD
   • (SPH)
   • Zeitschrittverfahren
   • sehr kurz: Thermostate, Ensembles (mit Einschränkungen, also welche gibt es, was brauchen sie, ...)
3. short range algorithms
   • LC, VL + force calculation
   • shortly mention long-range: direct/Barnes-Hut/FMM
   • Domain partitioning, only sdd., mention halo+boundary particles!, more details in MN-BB-LB
4. HPC-Architectures
   • single node: shared-memory, but often NUMA
   • multi-node: distributed memory
     • also with heterogeneous architectures
   • used clusters: MAC Cluster / Linux Cluster / SuperMUC 1/2 / SuperMIC / Hazel Hen -- eventually later, before results?
5. Overview of this thesis / Outline
   • previous implications of HPC in ls1-mardyn in next section
   • New opt. of multi node in section -> MNO + AT in MN
   • New opt. of node-level opt in -> Node-Level AT

Notes:

• einfach nur grundlegende Algorithmik + Domain decomp

⁠ls1-mardyn

• what it is, what it does
• container: LC
• traversal: c08, sliced, c04
• vectorization: intrinsics up to AVX512
• with context of HPC Architectures

⁠(Scenarios)

Scenario Descriptions

1. artificial scenarios
   • homogeneous particle distribution
   • mksfera
   • mktcts
   • other
2. realistic scenarios
   • Expl. liquid
   • droplet coalescence
   • spinodal decomposition
   • toffifee?

Note:

• die nicht hier einführen, sondern wenn sie gebaucht werden, also später.

⁠Multi-Node Optimizations

⁠Building Blocks

1. MPI Overview
   • P2P
   • Collective
   • Non-Blocking
2. Load Balancing Overview
   • different algorithms (kdd, diffusive, space-filling curves, ...)
   • load estimation
     • Überleitung: this works good for, e.g., stencil based algorithms, but for MD load estimation/calculation is also needed
     • OPEN: how to handle related work here? in this section, or the next?
3. Zonal Methods
   • NT, MP, ...

⁠Related Work

1. More detailed view at Multi-Node optimizations of other software
   • lammps
   • Gromacs
   • maybe other software?
   • FDPS?
     • multi-section method, <-- maybe better to introduce later, i.e., at multi-node AT...
2. Sachen von vorherigen Dissertationen am Lehrstuhl
3. Sachen von Steffen Hirschmann

⁠Implementation and Results

1. load estimation and k-d decomposition in ls1 mardyn
   • Methods:
     • heterogeneous domain
     • heterogeneous arch -- Motivation: MAC cluster and other trends (new Jülich cluster?)
     • vectorization tuner
     • backward calculation (speed from measurements, matrix inverse, ...)
     • hetero + vec tuner
   • results:
     • broad comparison of the different algorithms, clusters and scenarios.
2. Neutral Territory/Zonal Methods
   • Implementation:
     • ZonalMethod classes: Specifying different import/export regions
       • MP: needs to be cell-based!
     • Neighbor Acquirer: Getting Neighbors from arbitrary Zonal Methods
     • Matching Traversals:
       • FS, ES, HS: Mainly c08
       • MP: own traversal
       • NT: own traversal
   • Results:
     • Scaling results, mostly important for large scale homogeneous experiments.
3. MPI optimizations
   • overlapping p2p
   • overlapping globals (evtl. physik angucken)
   • and new WR -> how much was overlapping useful here?

⁠Improving Node-Level Performance using Auto-Tuning

• with context of ls1 mardyn!

⁠Motivation

• during WR: sometimes sliced, sometimes c08 better
• idea:
  • can we extend this using auto-tuning
  • move force calculation of ls1 mardyn outside and make it tunable

⁠Building Blocks

• Short overview auto-tuning
• auto-vectorization?
• algorithm-selection geht
  • nachbarn
  • parallelisierung
  • datenstrukturen
• weiteres?

Notes:

• technisch und algorithmisch
• maybe merge this with related work, for VerletClusterLists, ...

⁠Related Work

1. What do other codes use for better node-level performance:
   • lammps:
     • verlet-lists
     • vectorization (by intel), cuda, kokkos, evtl. andere
   • Gromacs:
     • Verlet-Cluster lists
     • cuda
   • FDPS:
     • somewhat similar, but different!
     • node-level: internally represents an octree
2. Which software uses which containers
3. Lammps: Kokkos?
4. GPU in Gromacs

⁠Implementation and Results

1. General
   • Grundidee / Motivation, warum nützlich?
2. AutoPas library
   • short outside overview
   • inner workings
     • software design
3. verschiedene Optionen
   • Container
     • DS
     • LC, wenige Traversals
     • VL
     • VCL, cuda
   • Traversals
   • und andere ---
4. Auto-Tuning, short overview.
   • Full-Search
   • Full-SearchMPI
5. Node-Level results

⁠Enabling Auto-Tuning on the Multi-Node Level

mainly paper from 2020 <-- CITE THIS!

1. AutoPas + ls1 + LB
   • why different LB scheme?

⁠Summary and Outlook

⁠Summary

⁠Outlook

• AutoPas
  • potentially update to kokkos for performance portability
  • better auto-tuning strategies in development by Fabio
  • more tuning options
  • multi-centered molecules support
  • c++20 concepts for better error handling, ...
  • lammps integration for AutoPas is in the work
• ls1-mardyn
  • more load balancing libraries with autopas
  • adapt everything to AutoPas

⁠Appendix

Potentially:

• CI via Jenkins
• Funny Bug Collection: intel + rundungsfehler

⁠Notes

• welche building blocks brauche ich --> intro
• wofür gibts schon related work
• was gibt es für vorraussetzungen?
• szenarios eher später, also erst vor den ergebnissen
• kapitelnamen anpassen, eventuell ohne Softwarenamen?
• software und algorithmen eng verzahnen, oder, alternativ, erst algo, dann software?
• wo sehe ich meine Hauptbeiträge?
  • im wesentlichen bei ls1? -> eng verzahnen
  • neuer algorithmus -> nicht eng verzahnen
  • => mal aufschreiben, wie ich das sehe