SHF: Medium: Intelligent and Efficient Data Movement for Multicore Systems

拥有数百个内核的多核芯片可能很快就会上市。目前的趋势表明，核心将相对简单，片上存储器将被分区为每个核心的高速缓存，并且每个高速缓存将相对较小。此外，芯片将继续受到管脚限制，因此DRAM带宽不会随着核的数量而扩展。由于片上缓存的总量很大，数据密集型并行软件有可能在多核处理器上运行得很好，但需要有效地利用分布式片上缓存。危险在于，许多内核使用的数据将被复制到许多内核的缓存中，从而减少了可缓存的不同数据的总量；一些内核可能会导致尝试访问超出其缓存容量的DRAM负载，而其他内核具有空闲的缓存空间。随着每个DRAM接口的内核数量和芯片上缓存之间的延迟增加，软件管理好分布式缓存非常重要。一种涉及架构师和系统研究人员的协作方法，以扩展存储器接口以提供对数据放置的控制并实现片上高效的数据移动，从而提供了整体解决方案。具体地说，这项工作提出了数据移动控制(DMC)接口，其中包括支持高速缓存到高速缓存的传输、对多个高速缓存线的批处理、向靠近保存特定数据项的高速缓存的核心发送消息以及数据监控。该建议还包括通过互连来路由数据的技术，这将为DMC接口提供高性能。如果成功，DMC将帮助数据密集型应用程序避免内存瓶颈，并使它们能够在未来的多核处理器上良好运行。全系统模拟器将允许其他研究人员探索智能和高效的数据管理，并支持麻省理工学院计算机体系结构研究生课程的研究项目，教育学生多核计算的挑战。

Multicore chips with hundreds of cores will likely be available soon. Current trends suggest that cores will be relatively simple, that on-chip memory will be partitioned into per-core caches, and that each cache will be relatively small. Furthermore, chips will continue to be pin-limited and therefore DRAM bandwidth won't scale with the number of cores.Parallel software that is data intensive has the potential to run well on multicore processors due to the large total amount of on-chip cache, but requires effective use of the distributed on-chip caches. The dangers are that data used by many cores will be duplicated in many cores' caches, decreasing the total amount of distinct data that can be cached; and that some cores may incur DRAM loads trying to access more data than fits in their caches, while other cores have spare cache space.As the number of cores per DRAM interface and the latencies between caches on a chip increase, it is important for software to manage the distributed caches well. A collaborative approach that involves architects and systems researchers to extend the memory interface to provide control over data placement and enable on-chip efficient data movement provides a holistic solution. Specifically, this work proposes data-movement control (DMC) interface, which includes support for cache-to-cache transfers, batching of several cache lines, sending a message to a core that is close to a cache that holds a particular data item, and data monitoring. The proposal also includes techniques to route data over the interconnect that will provide high performance for the DMC interface.If successful, DMC will help data-intensive applications avoid memory bottlenecks and allow them to run well on future multicore processors. A full-system simulator will allow other researchers to explore intelligent and efficient data management and enable research projects in MIT's graduate computer architecture class, educating students about the challenges of multicore computing.