In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Aaron Welch

will defend his dissertation proposal

Addressing Memory Allocation and Management in Exascale and Beyond

Abstract

In the push toward exascale and beyond, it has become clear that the technology and patterns that have served us well in the past are quickly becoming unreasonable to scale in the same manner as power requirements explode and we are unable to achieve the desired performance improvements within affordable power envelopes. This necessitates shifts in our approach to many levels of the design for both hardware as well as software. Among these shifts is a novel diversification of the memory hierarchy, with an increasingly complex assortment of memory devices and storage/access qualities. This introduces the difficult problem of addressing how to optimally make use of these hierarchies as well as how to efficiently implement such strategies. A number of different programming languages and libraries have been making strides to support the paradigms these new systems represent. However, this has led to several separate efforts for solving largely the same problems with a variety of different and incompatible abstractions. While there is value in having multiple approaches to software design that possess different characteristics when they are mutually exclusive, it is likely that what these competing efforts hope to create is functionally similar enough at a lower level that much of the underlying framework could benefit from a shared subsystem that is capable of servicing the needs of each of the varying models. As such, we will look at many of these different environments and their current or future feature changes as it pertains to complex memory hierarchies, and then attempt to reconcile their differences in the design of an allocator that can be used by all of them.

Date: Friday, January 18, 2019
Time: 9:45 AM
Place: PGH 501D
Advisors: Dr. Edgar Gabriel

Faculty, students, and the general public are invited.