[Defense] Extending the OpenSHMEM Communication Model Into the Exascale Era
Monday, April 17, 2023
4:30 pm - 6:30 pm
In
Partial
Fulfillment
of
the
Requirements
for
the
Degree
of
Doctor
of
Philosophy
Aaron
Welch
will
defend
his
dissertation
Extending
the
OpenSHMEM
Communication
Model
Into
the
Exascale
Era
Abstract
OpenSHMEM is an API for the Partitioned Global Address Space programming model that is used in many application domains for the simplicity of its memory abstractions and proximity to the low level Remote Direct Memory Access network performance features it provides. However, its origin as a thin interface to the underlying hardware aged along with the system architecture that gave birth to it, and is no longer as naturally suited to express the features of modern and emerging architectural trends. Of particular concern are difficulties with scaling small and irregular remote memory accesses on modern interconnects along with the increasingly complex memory hierarchies they access. This work investigates what kinds of extensions to its design may be required to better adapt to the diversity of future computing systems whilst remaining as close as possible to the traditional OpenSHMEM model and retaining its low overhead. To that end, we will first look at the construction of a highly performant reference implementation capable of directly leveraging latent capabilities of network technology via a performance portable low-level communication middleware in order to drive public interest and development. Then, we will look at a series of extensions to the OpenSHMEM model aimed at increasing its flexibility and maximising its performance, notably to increase memory and inter-process locality for exploiting resource affinity, and message aggregation abilities to achieve optimal use of network resources. Finally, we will present a tool intended to use high level patterns for identifying regions of applications that could benefit from these new extensions as well as analyse the resulting application performance behaviours.
Monday,
April
17,
2023
4:30PM
-
6:30PM
CT
Online
via
Dr. Jaspal Subhlok, Faculty Advisor
Faculty, students and the general public are invited.
