Research On Stream Program Virtual Machine For Multi-core Processor With Distributed Memory

With multi-core processors having become mainstream products and industry standards, how to simplify the programming style by ignoring low-level details of processor, while making the best use of multi-core processor's parallel resources, is a hot and hard issue in parallel programming. The stream programming paradigm offers a promising approach by expressing parallelism explicitly, that is widely used in data- and computation- intensive domains such as multi-media processing, science computing and etc. With the extensive concern on stream programming model, many chip manufacturers and research institutions have developed different kinds of multi-core architectures to support the stream applications, but the software systems are only suitable for their own platforms, thus it's hard for applications to transplant among different platforms. To deal with this problem, a systematic study of stream applications'cross-platform is needed.By adding data flow model to the Brook programming model, a new stream programming model DFBrook is designed to support data-level and task-level parallelism. Programmers can invoke the data-flow execution with specific program code modules embedded in the regular C code. To make DFBrook applications easily portable among different processors, a Stream program Virtual Machine (SVM) is created to optimize stream applications'excution. SVM abstracts key components of the target machines such as memory hierarchy, processor architecture and direct memory access engines, to form an abstract machine model. The SVM also defines a set of API that each machine implements mapping stream program onto multicore processors in the way of Compiler-Assisted Message-Driven Dynamic Scheduling.A set of experiments is designed with CELL processor and Godson_T processor as hardware environment, and common algorithms in digital media domain as benchmarks, to evaluate SVM's performance. Experimental results demonstrate that the SVM is scalable with processor cores and nearly reaches linear speedup on processors while keeping high load balance efficiency and low communication overhead.