Enzo-P / Cello: scalable adaptive mesh refinement for astrophysics and cosmology

• Authors: James Bordner;Michael L. Norman
• Venue: Proceedings of the Extreme Scaling Workshop
• Year: 2012

Cello is a highly-scalable object-oriented adaptive mesh refinement (AMR) framework currently under development. While Cello is expected to be usable within multiple scientific problem domains, we specifically target specialized requirements of astrophysics and cosmology applications. Development of Cello is funded by the National Science Foundation (PHY-1104819, AST-0808184) The Cello project grew from the need to address scalability in the parallel AMR astrophysics and cosmology application Enzo [12, 41, 45]. Enzo has a long proven track record in producing new scientific results [1, 30, 36, 47]; however, its AMR design and implementation have several known scaling issues that are difficult to address without costly and invasive changes to the code. This has made it progressively more difficult for Enzo to take full advantage of the compute power of current high-end HPC platforms. While development still continues on improving Enzo's scalability, we are also implementing Enzo's physics capabilities using the Cello scalable AMR framework. The resulting "petascale" version of Enzo is called Enzo-P. In this paper we elaborate on the known scaling issues in Enzo, and how we plan to address these and other scaling issues in Cello through a combination of existing and novel approaches. Two of the more fundamental changes are to incorporate process virtualization and data-driven execution by using Charm++ [34], and to use a new variant of the "forest-of-octrees" approach for the AMR infrastructure.