CSR: Medium: A Computing Cloud for Graphical Simulation

Stanford University, Stanford CA

Investigators

Abstract

Today, many graphical simulations run on a single powerful server or a small cluster of high-performance, high-cost nodes. This research aims to answer the question -- is it possible to run graphical simulations in the computational cloud? -- by designing and implementing Nimbus, a software for graphical simulation in the computing cloud. The goal is to be able to run large, complex simulations using on-demand cloud computing systems. Nimbus supports PhysBAM, an open-source graphical simulation package developed and maintained by Principal Investigator Fedkiw. The project will collaborate with existing PhysBAM users to support the Nimbus software for broader use and adoption. Nimbus focuses on three important principles to support graphical simulations running on hundreds to thousands of cloud servers. First is decoupling data access and layout. Nimbus represents data in three layers: program, logical, and physical. These layers separate the units which a program operates on (program) from the units which the Nimbus software manages and transfers (logical) from how they are laid out in actual computer memory (physical). Second is non-uniform, geometry-aware data placement. Nimbus uses the fact that simulations have a basic underlying geometry to intelligently place data and computation. This geometry is explicit in the Nimbus software, which knows that nearby regions of the simulation should be placed on nearby computers. Third is dynamic assignment and load balancing: Graphical simulations today divide the simulation volume equally across computers, despite the fact that some regions require much more computation than others. Nimbus divides a simulation into a larger number of smaller partitions, which it dynamically assigns and moves as load changes to reduce running time while considering inter-partition communication. These three principles allow Nimbus to provide tremendous flexibility. The system breaks a simulation into small pieces that a controller computer sends to worker computers to compute. These worker computers decide when to schedule these simulation pieces and how to assign processors to different pieces. The runtime automatically moves data in the most efficient manner possible as needed, compressing data and replicating it when having multiple copies for different pieces increases performance. Discovering how these applications can be run on modern data center computing systems will help bring arithmetically intensive scientific computing to the cloud. As Exascale and other supercomputing efforts gain momentum, their scale will need to deal with the same issues cloud systems have been tackling for the past decade, stragglers, failures, and heterogeneity. By focusing on one particular compelling application, this work will establish an intellectual framework for future, broader efforts.