2014 Poster Sessions : Nimbus: Runtime System for Graphical and Scientific Simulations over Cloud

Student Name : Chinmayee Shah, Hang Qu, Omid Mashyaekhi
Advisor : Philip Levis
Research Areas: Computer Systems
Even today, after the emergence of commercial clouds, graphical simulations and scientific computations are restricted to super-computers or small, customized, high performance clusters. These simulations tend to run for several days or weeks and assume reliable and uniform resources over each host. Current techniques for distributing these simulations do not provide mechanisms to deal with stragglers and faults. On the other hand, existing cloud frameworks like MapReduce do not provide any notion of geometric locality between data and computation, making them a poor fit for graphical simulations. We will present Nimbus, a runtime framework that uses information about geometric locality, and provides mechanisms to deal with non-uniform computation and resources. Nimbus provides an abstraction that hides the complexity of dealing with actual physical data, tasks and their placement over the cloud, from the application writer. To evaluate Nimbus in real world applications, we have ported an incompressible fluid simulation from the PhysBAM library into the Nimbus framework. PhysBAM is a physics based library for graphical simulations developed at Stanford University, and is widely used by various groups and in film industry.

Chinmayee Shah is a third year PhD student at Stanford University, advised by Prof. Philip Levis. She is interested in abstractions for graphics simulations, to enable them to run on the cloud. She is also interested in distributed dynamic data structures and high performance for graphics applications.

Omid Mashayekhi is a PhD candidate at stanford University. He received his Master's degree from Stanford University in 2013, and before that Bachelor of Science degree from Sharif University in 2011. Currently he is advised by Professor Levis and working on new abstractions and frameworks to run computationally expensive graphical simulations over large clusters of commodity machines. He is interested in distributed computing, parallel processing scheduling and networking systems.

Hang Qu is a second year PhD candidate at Stanford University, advised by Prof. Philip Levis. He is generally interested in running graphics simulation in the cloud.