CS267 / Eng233 / IDS267 Applications of Parallel Computers
Lecture 4: Large-Scale Parallel Architecture
Abstract
This lecture continues our tour through the evolution of parallel architectures.
First we pick up the thread of message passing architectures. This takes us on
a little diversion through networks, before we come back to understand that
modern large scale machines are quite flat, in spite of the interconenction
topology. The message passing model and the hardware are separated by a
software layer that implements a protocol using more primitive network transactions
to realize the semantics of the message passing model. We look at some examples
and compare performance. These underlying primitives are the basis for
large-scale shared address space machines as well, however, the protocols
are implemented directly in hardware. Finally, we pick up the data parallel
SIMD thread, which brings us to the full convergence.