Towards Distributed Petascale Computing

TL;DR

This paper discusses the need for hybrid, distributed computing models to simulate complex multi-scale, multi-physics systems like the galaxy, requiring Petaflop/s scale computational power.

Contribution

It proposes a framework for distributed Petascale computing tailored for multi-scale multi-physics simulations, exemplified by galaxy modeling.

Findings

Estimate of compute requirements for galaxy simulation

Highlighting the importance of hybrid distributed models

Discussion on data proximity in large-scale simulations

Abstract

In this chapter we will argue that studying such multi-scale multi-science systems gives rise to inherently hybrid models containing many different algorithms best serviced by different types of computing environments (ranging from massively parallel computers, via large-scale special purpose machines to clusters of PC's) whose total integrated computing capacity can easily reach the PFlop/s scale. Such hybrid models, in combination with the by now inherently distributed nature of the data on which the models `feed' suggest a distributed computing model, where parts of the multi-scale multi-science model are executed on the most suitable computing environment, and/or where the computations are carried out close to the required data (i.e. bring the computations to the data instead of the other way around). We presents an estimate for the compute requirements to simulate the Galaxy as a typical example of a multi-scale multi-physics application, requiring distributed Petaflop/s computational power.