Michael Stumm: Alumni

Ph.D. Alumni: Robin Grindley

Reference:

Robin Grindley
The NUMAchine Multiprocessor: Design and Analysis
Ph.D. Thesis, Department of Electrical and Computer Engineering, University of Toronto, Toronto, Canada, 1990.
Primary supervisor: Prof. Zvonko G. Vranesic

Supervisor(s):

Zvonko G. Vranesic (primary)
Michael Stumm

Download Thesis:

PDF

Abstract:

This dissertation considers the design and analysis of NUMAchine: a distributed, sharedmemory multiprocessor. The architecture and design process leading to a working 48-processor prototype are described in detail. Analysis of the system is based on a cycle-accurate, execution-driven simulator developed as part of the thesis. An exploration of the design space is also undertaken to provide some intuition as to possible future enhancements to the architecture.

Shared-memory multiprocessors and parallel processing are becoming increasingly common not only in the scientific domain, but also as a replacement for mainframes in the field of large-scale enterprise computing. The shared-memory programming paradigm provides an intuitive view of memory as a globally shared resource among all processors. This is more familiar to programmers of uniprocessors than the alternative, message-passing. The distribution of memory across the system leads to Non-Uniform Memory Access times (NUMA), since processors have fast access to local memory and slower access to remote memories across the system network.The architecture contains features which attempt to hide or reduce the effects of this non-uniformity.

NUMAchine provides cache coherence in hardware, making it an instance of the general class of multiprocessor architectures called CC-NUMA (for cache-coherent NUMA). Thesystem network in NUMAchine consists of a hierarchy of rings. We show how certain properties of rings allow for an efficient cache coherence scheme with reduced overheads in comparison to other CC-NUMA architectures. We use the simulator, which we developed as part of this project, to explore the NUMAchine design space in an attempt to discover how changes in various aspects of the architecture affect overall performance.

Keywords:

Computer architecture, shared memory multiprocessors, cache consistency, NUMA multiprocessors

BibTeX:

@phdthesis(Grindley-PhD99,
    author = {Robin Grindley},
    title = {The NUMAchine Multiprocessor: Design and Analysis},
    school = {Department of Electrical and Computer Engineering, University of Toronto},
    address = {Toronto, Canada},
    supervisors = {Zvonko G. Vranesic (primary), Michael Stumm},
    year = {1990},
    keywords = {Computer architecture, shared memory multiprocessors, cache consistency, NUMA multiprocessors}
)