Multi-cluster parallel job submission: Experiences with Monte Carlo simulations for computational finance on Grid5000

Ian Stokes-Rees (1) , Françoise Baude (1) , Viet Dung Doan (1) , Mireille Bossy (2)

1 OASIS - Active objects, semantics, Internet and security
2 TOSCA - Simuler et calibrer des modèles stochastiques

As experience with independent and embarassingly parallel computations in a grid environment mature, it has become possible to explore parallel computing on grids with higher levels of inter-task communication. The PicsouGrid project applies grid computing concepts to computational finance, aiming to leverage heterogeneous resources for both time critical and high volume computations. It utilizes the ProActive Java distributed computing library to parallelize and distribute Monte Carlo option pricing simulations, concurrently utilizing 10^2 −10^3 workers. PicsouGrid has been deployed on various grid systems to evaluate its scalability and performance. Issues arising from the heterogeneity and layering of grid infrastructures are addressed via an abstract process model which is applied at each layer. Timings of both the algorithms and the grid infrastructures are carefully measured to provide better insight into the behaviour and utilization of computational grids for this important class of parallel simulation.

Mots clés: computational grid; computational finance; grid performance

Domaines: Calcul parallèle, distribué et partagé [cs.DC]

Ian Stokes-Rees, Françoise Baude, Viet Dung Doan, Mireille Bossy. Multi-cluster parallel job submission: Experiences with Monte Carlo simulations for computational finance on Grid5000. 2007. ⟨inria-00173360⟩