Toward Large-Scale Shared Memory Multiprocessing

Abstract

We are currently investigating two different approaches to scalable shared memory: Munin, a distributed shared memory (DSM) system implemented entirely in software, and Willow, a true shared memory multiprocessor with extensive hardware support for scalability. Munin allows parallel programs written for shared memory multiprocessors to be executed efficiently on distributed memory multiprocessors. Unlike existing DSM systems, which only provide a single mechanism for memory consistency, Munin provides multiple consistency protocols, matching protocol to data object based on the expected pattern of accesses to that object. We call this approach type-specific coherence. Munin also employs a relaxed consistency model to mask network latency and to minimize the number of messages required for keeping memory consistent. Willow is intended to be a true shared memory multiprocessor, providing memory capacity and performance capable of supporting over a thousand commercial microprocessors. These processors are arranged in cluster fashion, with a multi-level cache, I/O, synchronization, and memory hierarchy. Willow is distinguished from other shared memory multiprocessors by a layered memory organization that significantly reduces the impact of inclusion on the cache hierarchy and that exploits locality gradients. Willow also provides support for adaptive cache coherence, an approach similar to Munin’s type-specific coherence, whereby the consistency protocol used to manage each cache line is selected based on the expected or observed access behavior for the data stored in that line. Implementation of Munin is in progress; we are still designing Willow.