Protection and Translation in Multimodel Multiprocessors: The MIP FUGU Workstation

该项目正在扩展AleWare多处理器，以探索未来可伸缩多处理器的两个方面。一方面是保护。对硬件资源的访问需要保护，尽可能少的软件开销。例如，当消息发送者和接收者是同一应用程序的一部分时，它们应该能够直接访问网络硬件。另一个方面是地址转换，这样每个应用程序都可以有自己的地址空间。但是，执行应用程序的所有处理器必须具有相同的地址空间。为了确保这一点，内存映射硬件被放置在多处理器的每个节点上，并且必须保持一致。河豚系统提供三种处理器间通信模式。共享内存由节点之间的硬件合成消息提供，以维护缓存一致性。通过对每个节点地址空间中的I/O寄存器执行加载和存储，可以直接发送短消息。较长的消息由独立于处理器的DMA引擎管理。通过硬件增强来提供对共享存储器和短消息的转换和保护，这些硬件增强维护转换后备缓冲器中的一致性，并确保消息仅由它们被寻址的进程接收。对于长消息，硬件只提供对地址转换的最低限度的支持，因为这种转换的需要可能很少。任何需要的翻译都将由软件提供。

This project is extending the Alewife multiprocessor to explore two aspects of future scalable multiprocessors. One aspect is protection. Protection is required on accesses to hardware resources, with as little software overhead as possible. For example, when a message sender and receiver are part of the same application, they should be able to directly access network hardware. The other aspect is address translation, so that each application can have its own address space. All processors executing an application must have the same address space, however. To ensure this, memory mapping hardware is placed at each node of the multiprocessor, and must be kept coherent. The FUGU system provides three modes of interprocessor communication. Shared memory is provided by hardware synthesized messages among nodes, which maintain cache coherence. Short messages can be sent directly by performing loads and stores to I/O registers that are in each node's address space. Longer messages are managed by a DMA engine that is separate from the processor. Translation and protection for shared memory and for short messages are provided by hardware enhancements that maintain coherence in the translation lookaside buffers and ensure that messages are received only by the processes that they are addressed to. For long messages, only minimal support for address translation is provided by the hardware since the need for such translation is likely to be rare. Any needed translation will be provided by software.