Computer Science ETDs

Publication Date

Fall 12-13-2020

Abstract

There exists a resurgence of interest in `smart' network interfaces that can operate on data as it flows through a network. However, while smart capabilities have been expanding, what they can do for high-performance computing (HPC) is not well-understood. In this work, we advance our understanding of the capabilities and contributions of smart network interfaces to HPC. First, we show current offloaded message demultiplexing can mitigate (but not eliminate) overheads incurred by multithreaded communication. Second, we demonstrate current offloaded capabilities can be leveraged to provide Turing complete program execution on the interface. We elaborate with a framework for offloading arbitrary compute kernels to the NIC: In-Network Compute Assistance (INCA). We show INCA can accelerate host applications by offloading components to the network. Moreover, INCA supports the offloading of autonomous machine learning kernels for predicting network properties, and by doing so, takes a significant first step towards realizing intelligent, adaptive networks.

Language

English

Keywords

High-performance computing, networks, machine learning

Document Type

Thesis

Degree Name

Computer Science

Level of Degree

Doctoral

Department Name

Department of Computer Science

First Committee Member (Chair)

Trilce Estrada

Second Committee Member

Dorian Arnold

Third Committee Member

Ryan E. Grant

Fourth Committee Member

Jinho D. Chen

Project Sponsors

Sandia National Laboratories

Download

Share

COinS