| Summary: | Traditional scientific applications such as Computational Fluid Dynamics,
Partial Differential Equations based numerical methods (like Finite Difference
Methods, Finite Element Methods) achieve sufficient efficiency on state of the
art high performance computing systems and have been widely studied / implemented using
conventional programming models. For emerging application domains such as Graph applications
scalability and efficiency is significantly constrained by the conventional systems and
their supporting programming models.
Furthermore technology trends like multicore, manycore,
heterogeneous system architectures are introducing new challenges and possibilities.
Emerging technologies are requiring a rethinking of approaches to more effectively
expose the underlying parallelism to the applications and the end-users.
This thesis explores the space of effective parallel execution of ephemeral graphs
that are dynamically generated. The standard particle based simulation,
solved using the Barnes-Hut algorithm is chosen to exemplify the
dynamic workloads.
In this thesis the workloads are expressed using sequential execution semantics,
a conventional parallel programming model - shared memory semantics and semantics
of an innovative execution model designed for efficient scalable performance towards
Exascale computing called ParalleX. The main outcomes of this research are parallel
processing of dynamic ephemeral workloads, enabling dynamic load balancing during runtime, and
using advanced semantics for exposing parallelism in scaling constrained applications.
|
|---|
