Research Initiation Award: Incorporating Fault Tolerance in the Multiscalar Processor

9410706富兰克林多标量处理器是最近提出的一种利用指令级并行的执行范例。它是使用环型网络以逻辑分离和物理分散的方式连接在一起的顺序处理器的集合。其核心思想是将指令流划分为指令块，然后通过使用多个处理元件执行这些块来利用并行性。多标量处理机在学术界和工业界都引起了广泛的兴趣，似乎有巨大的潜力成为约2000指令级并行处理机的处理范例。本研究的目的是开发一种高性能的容错多标量处理器，并推动容错技术在通用处理器中的应用。将容错结合到多标量处理器中的主要方式是通过软件的运行时复制以及通过在处理器的多个执行单元上执行副本来实现容错，这些执行单元是出于性能原因而提供的。两个重新执行方案(都使用相同的复制技术)正在调查中。已经进行了初步的仿真研究，以了解由于容错而导致的性能损失。这些初步结果表明，时间冗余技术是在多标量处理器中提供容错的一种很好的选择。***

9410706 Franklin The multiscalar processor is a recently proposed execution paradigm for exploiting instruction-level parallelism. It is a collection of sequential processors that are connected together in a logically decoupled and physically decentralized manner using ring-type network. The central idea is to divide an instruction stream into blocks of instructions, and then exploit parallelism by executing these blocks using multiple processing elements. The multiscalar processor has generated wide interest, both in the academia and the industry, and seems to have tremendous potential to become the processing paradigm of choice for a circa 2000 intruction level parallel processor. The purpose of this research is to develop a high performance fault tolerant multiscalar processor and to provide an impetus to the incorporation of fault tolerance in general purpose processors. The primary manner fault tolerance will be incorporated in the multiscalar processor is by run-time replication of software and by executing the replicas on the processor's multiple execution units, which are provided for performance reasons. Two re-execution schemes (both using the same replication technique) are under investigation. Preliminary simulation studies have been conducted to get an idea of the performance loss incurred because of fault tolerance. These preliminary results indicate that time-redundant techniques are an excellent choice for providing fault tolerance in the multiscalar processor. ***