GPUs are parallel devices that are able to run thousands of
independent threads concurrently. Traditional GPU programs are
data-parallel, requiring little to no communication,
i.e. synchronisation, between threads. However, classical concurrency
in the context of CPUs often exploits synchronisation idioms that are
not supported on GPUs. By studying such idioms on GPUs, with an aim to
facilitate them in a portable way, a wider and more generic space of
GPU applications can be made possible.

While the breadth of this thesis extends to many aspects of GPU
systems, the common thread throughout is the global barrier: an
execution barrier that synchronises all threads executing a GPU
application. The idea of such a barrier might seem straightforward,
however this investigation reveals many challenges and insights. In
particular, this thesis includes the following studies:

Execution models: while a general global barrier can deadlock due to
starvation on GPUs, it is shown that the scheduling guarantees of
current GPUs can be used to dynamically create an execution
environment that allows for a safe and portable global barrier
across a subset of the GPU threads.


Application optimisations: a set GPU optimisations are examined that
are tailored for graph applications, including one optimisation
enabled by the global barrier. It is shown that these optimisations
can provided substantial performance improvements, e.g. the barrier
optimisation achieves over a 10X speedup on AMD and Intel GPUs. The
performance portability of these optimisations is investigated, as
their utility varies across input, application, and architecture.

Multitasking: because many GPUs do not support preemption,
long-running GPU compute tasks (e.g. applications that use the
global barrier) may block other GPU functions, including graphics. A
simple cooperative multitasking scheme is proposed that allows
graphics tasks to meet their deadlines with reasonable overheads.