Satin: Simple and efficient Java-based grid programming. Scalable Computing: Practice and Experience

Abstra t. Grid programming environments need to be both portable and e ient to exploit the omputational power of dynami ally available resour es. In previous work, we have presented the divide-and- onquer based Satin model for parallel omputing on lustered wide-area systems. In this paper, we present the Satin implementation on top of our new Ibis platform whi h ombines Java's write on e, run everywhere with e ient ommuni ation between JVMs. We evaluate Satin/Ibis on the testbed of the EU-funded GridLab proje t, showing that Satin's load-balan ing algorithm automati ally adapts both to heterogeneous pro essor speeds and varying network performan e, resulting in e ient utilization of the omputing resour es. Our results show that when the wide-area links suer from ongestion, Satin's load-balan ing algorithm an still a hieve around 80 % e ien y, while an algorithm that is not grid aware drops to 26 % or less. Key words. Satin, Ibis, divide-and- onquer, load balan ing, distributed super omputing. 1. Introdu tion. In omputational grids, appli ations need to simultaneously tap the omputational power of multiple, dynami ally available sites. The rux of designing grid programming environments stems ex-a tly from the dynami availability of ompute y les: grid programming environments need to be both portable to run on as many sites as possible, and they need to be exible to ope with dierent network proto ols and dynami ally hanging groups of heterogeneous ompute nodes. Existing programming environments are either portable and exible (Jini, Java RMI), or they are highl