CAREER: Sparks: Coherence as an Abstract Type

软件分布式共享内存（DSM）系统向工作站网络和分布式内存机器（如Paragon、CM-5或SP-2）提供共享内存的抽象。不幸的是，在这两种环境中，全局操作的延迟都比紧耦合的多处理器要昂贵好几个数量级。结果是，当前的dsm仅在有限的应用类别中表现良好。Sparks项目旨在通过提供一个协议构建库来解决这一限制，该库将允许DSM性能提高到紧密耦合多处理器的几个百分点以内。Sparks的抽象允许清晰而系统地探索高级同步操作的设计空间，而不是以一种特别的方式。重点更多地放在创建和研究使各种各样的优化成为可能的抽象，而不是单个优化本身。然而，通过Sparks库创建的同步类型所带来的性能提升将被彻底量化。Sparks接口将在CVM（一个新的分布式共享内存系统）中实现。除了为Sparks提供支持服务外，CVM还实现了新颖的一致性协议和多线程支持。＊＊＊

Software distributed shared memory (DSM) systems provide the abstraction of shared memory to networks of workstations and distributed memory machines, such as the Paragon, CM-5, or SP-2. Unfortunately, the latencies for global operations in either environment are several orders of magnitude more expensive than on tightly-coupled multiprocessors. The result is that current DSMs perform well for only a restricted class of applications. The Sparks project is designed to address this restriction by providing a protocol construction library that will allow DSM performance to be improved to within a few percent of tightly-coupled multiprocessors. Sparks' abstractions allow the design space of high-level synchronization operations to be explored cleanly and systematically, rather than in an ad hoc fashion. The emphasis is more on creating and investigating the abstractions that make a broad variety of optimizations possible, rather than on the individual optimizations themselves. However, the performance gains allowed by the synchronization types created via the Sparks library will be thoroughly quantified. The Sparks interface will be implemented in CVM, a new distributed shared memory system. In addition to providing support services to Sparks, CVM implements novel coherence protocols and multi- threading support. ***