Improving the Performance of Shared Virtual Memory on System Area Networks (Thesis)

Abstract:

As clusters of workstations, uniprocessor or symmetric multiprocessors
(SMPs), become important platforms for parallel computing, there is
increasing research interest in supporting the attractive, shared
address space programming model across them in software. The reason is
that it may provide successful low-cost, high-performance
alternatives to both tightly-coupled, hardware-coherent distributed
shared memory machines and to scalable servers. In both these cases,
the clusters are formed with off-the-self, high-end PCs or
workstations and system area networks that track technologies
well. Given that a shared memory abstraction is an attractive
programming model for this architecture, there has been a lot of
research in fast communication on clusters connected with system area
networks and in protocols for supporting software shared memory across
them. However, the end performance of applications that were written
for the more proven hardware-coherent shared memory is still not very
good on these systems.

This dissertation is about improving the performance of shared virtual
memory on clusters that are connected with system area networks. In
this architecture, three major layers of software (and hardware) stand
between the end user and performance, each with its own functionality
and performance characteristics. The lowest layer is the
communication layer: the communication hardware and the low level
software libraries that provide basic messaging facilities. Next is
the protocol layer that provides the programming model to the
parallel application programmer. Finally, above the programming model
or protocol layer runs the application itself. This
dissertation improves the performance of shared virtual memory on
clusters by studying and improving each of this layers and some of
their interactions.