Experiences with a computational fluid dynamics code on clouds, grids, and on-premise resources

Heterogeneity is often manifested in different forms even in environments that seem to be similar. In particular, clusters, grids, and infrastructure as a service (IaaS) clouds may appear straightforward to configure to be interchangeable. However, our experiences with mainstream parallel codes for solving complex computational fluid dynamics (CFD) problems demonstrate that secondary attributes – support software, interconnect type, availability, access, and cost – expose heterogeneous aspects that impact overall effectiveness of application execution as well as overall productivity. Our experiences provide preliminary insights into characterizing three different types of platforms to which users typically have access – local clusters, grids, and clouds – and show where the tradeoffs can be, in terms of deployment effort, actual and nominal costs, application performance, and availability (both in terms of resource size and time to gain access). The considered benchmark is the numerical solution of partial differential equations for a blood flow problem. For this application, we compare five platforms with respect to two metrics: time to completion and cost per simulation. We further introduce a third metric, the utility function of the simulation, and discuss different user profiles leading to different rankings of the tested platforms. Additionally, we propose a systematic approach to overcome heterogeneity introduced by software dependencies required for application execution. Our results suggest that IaaS clouds are a viable approach for intensive CFD simulations. Email address: {tpasser, jslawin, uvilla, avenez2, vss}@emory.edu (Tiziano Passerini, Jaroslaw Slawinski, Umberto Villa, Alessandro Veneziani, Vaidy Sunderam) Preprint submitted to Journal of Parallel and Distributed Computing October 16, 2012