In-Network Coherence Techniques for Scalable Shared Memory.

摘要

Moore's law has driven the continuous miniaturization of transistor size over the years. This has led to breakthroughs in the domain of computers (desktops, servers, supercomputers, etc.), as well as in the domain of mobile devices. Computers of today have abundant on-chip resources that can be exploited to perform increasingly complex tasks. Similarly, the computation capacity of mobile devices has grown manifold over the years. Today's high-end mobile devices have greater computational capability than the desktops available a few years ago. However, programming general-purpose computers as well as mobile devices is not easy. This dissertation seeks to address the parallel programming challenge existing in the domain of multi-core systems, as well as the distributed programming problem associated with mobile ad-hoc networks. Specifically, we look into providing scalable shared memory for both parallel and distributed mobile computing.;With technology scaling leading to an abundance of on-chip resources and uniprocessor designs providing diminishing returns, the computing industry has moved beyond single-core microprocessors and embraced the many-core wave. It is widely believed that shared memory programming is a key to addressing the parallel programming challenge. To implement shared memory, scalable cache coherence protocol implementations are necessary to allow fast sharing of data among various cores and drive the many-core revolution forward. In this dissertation, we take an in-network approach to address the many-core cache coherence problem. We tackle the ordering and bandwidth overhead problem of snoopy protocols in the interconnect and demonstrate how snoopy protocols, originally proposed for small-scale multiprocessors connected via ordered interconnects, present a viable cache coherence solution for many cores on a single chip connected via unordered interconnects.;To address the cache coherence ordering problem, we present In-network Snoop Ordering (INSO), in which coherence requests from a snoop-based protocol are inserted into the interconnect fabric and the network orders the requests in a distributed manner, creating a global ordering among requests. To address the broadcast bandwidth overhead of snoopy protocols, we propose embedding small in-network coherence filters (INCFs) inside on-chip routers that dynamically track sharing patterns among various cores and filter away redundant snoop requests to save interconnect bandwidth and power. This dissertation also addresses the problem of evaluating future many-core architecture proposals. We have implemented GARNET, a detailed and cycle-level on-chip interconnect model, inside the GEMS full-system simulator. GEMS, along with GARNET, provides a full-system performance and power evaluation framework.;Location-based services accessed by mobile devices are increasingly pervasive. Much of the data processed by such services are distributed near the locations where the processed results are needed. If the mobile devices could conduct some of this computation locally, they could gain three major advantages: (1) ease the bandwidth pressure on already overloaded access networks, (2) lead to quicker response times, and (3) improve battery life of mobile devices. To realize the above vision, there needs to be a way to easily program a collection of mobile devices as a whole, since a lot of programming complexities arise due to the mobile nature of such platforms.;In this dissertation, we present a programming abstraction layer, called the Consistent Shared Memory Layer (CSMlayer), that provides consistent shared memory objects to mobile devices. We argue that consistent shared memory is a key requirement in enabling an easy to program, distributed mobile platform. The key to realizing this vision is porting ideas of consistency and coherence from the multi-core domain to the wireless domain. We illustrate how applications that require data to remain reliable and be consistently accessed are now enabled over mobile nodes. This was impossible earlier.