Adaptive data parallelism for internet clients on heterogeneous platforms

Abstract

Today's Internet is long past static web pages filled with HTMLformatted text sprinkled with an occasional image or animation. We have entered an era of Rich Internet Applications executed locally on Internet clients such as web browsers: games, physics engines, image rendering, photo editing, etc. Yet today's languages used to program Internet clients have limited ability to tap to the computational capabilities of the underlying, often heterogeneous, platforms. In this paper we present how a Domain Specific Language (DSL) can be integrated into Action Script, one of the most popular scripting languages used to program Internet clients and a close cousin of JavaScript. We demonstrate how our DSL, called ASDP (Action Script Data Parallel), can be used to enable data parallelism for existing sequential programs. We also present a prototype of a system where data parallel workloads can be executed on either CPU or a GPU, with the runtime system transparently selecting the best processing unit, depending on the type of workload as well as the architecture and current load of the execution platform. We evaluate performance of our system on a variety of benchmarks, representing different types of workloads: physics, image processing, scientific computing and cryptography. Copyright © 2012 ACM.