Viability of the bulk synchronous parallel model for science on cloud

Abstract

In recent years, cloud computing has emerged as an alternative to classical HPC resources like supercomputers, computer clusters and grids. As cloud provides a convenient real-time access to dynamic computing resources, it appears ideal for solving large scale scientific computing problems in domains like physics, chemistry, astrophysics, climatology, etc. However, using a large number of cloud resources means that applications must be able to distribute the work between these resources in a fault tolerant manner as the commodity cloud resources are bound to fail in regular intervals. Frameworks like MapReduce can significantly simplify this by taking care of data distribution, synchronization, task parallelism, fault tolerance, etc. and only require the user to write the algorithms themselves. However, MapReduce model is not well suited for more complex iterative tasks that are commonly utilized in the field of scientific computing. This work investigates an alternative model, Bulk Synchronous Parallel (BSP), to find how suitable it and its current implementations are for running more complex scientific computing algorithms. We take a number of BSP implementations, adapt several typical scientific algorithms to them and compare the results to MPI based implementations. The goal of this work is to give an overview of the current BSP implementations, evaluate whether they are suitable for complex scientific algorithms in comparison to MPI and still provide the same advantages as MapReduce. © 2013 IEEE.