	Error rendering macro 'children' null

Abstract

This project will develop and deploy a novel instrument for accelerating deep learning research at the University of Illinois. The instrument will integrate the latest computing, storage, and interconnect technologies in a purpose-built shared-use system. This Instrument will deliver unprecedented performance levels for extreme data-intensive emerging fields of research with far-reaching impacts in many areas, such as computer vision, natural language processing, artificial intelligence, healthcare and education. The instrument development will be driven by the UofI deep learning community needs and will be carried out in collaboration with IBM and Nvidia. The instrument will serve as a focal point for the rapidly growing deep learning research community at UofI, enable expansion of several research programs, and contribute to STEM education and training.

Personnel

Co-PIs

Bill Gropp <wgropp@uiuc.edu>

Volodymyr Kindratenko <kindrtnk@illinois.edu>

Roy Campbell <rhc@illinois.edu>

Jian Peng <jianpeng@illinois.edu>

Development team

Volodymyr Kindratenko

Research engineer: TBD

Postdocs: 2 TBD

GRAs: TBD

Advisory Board

TBD

Campus collaborators

Participants Name	Department/ Affiliation	Title	Research Area
N. Hovakimyan	MechSE	Prof.	Agricultural Data Analysis
M. Hudson	Crop Sciences	Prof.	Bioinformatics
S. Sinha	CS	Prof.	Bioinformatics
E. Tajkhorshid	MCB	Prof	Bioinformatics
G. Robinson	IGB	Director of IGB	Bioinformatics
J. Peng	CS	Asst. Prof.	Bioinformatics
M. Do	ECE	Prof.	Computer Vision
R. Yeh	ECE	Grad. Student	Computer Vision
G. Allen	Astronomy/Education	Prof/Assoc. Dean Education	Cosmology & Astrophysics
R. Brunner	Astronomy	Prof.	Cosmology & Astrophysics
D. George	Astronomy	Grad. Student	Cosmology & Astrophysics
E. Escudero	NCSA	Post-doc	Cosmology & Astrophysics
M. Turk	Info. Sci.	Asst. Prof.	Cosmology & Astrophysics
R. Farivar	CS	Adj. Sr. Lecturer	Deep Learning
S. Wang	Geography	Prof	Earth and Environment
J. Sirignano	ISE	Asst. Prof.	Financial Engineering
Y. Fan	ECE	Grad. Student	Image Generation & Analysis
D. Forsyth	CS	Prof.	Image Generation & Analysis
T. Huang	ECE	Prof.	Image Generation & Analysis
G. Ko	ECE	Grad. Student	Image Generation & Analysis
J. Yu	ECE	Grad. Student	Image Generation & Analysis
J. Towns	NCSA	Executive Director	Information and Technology
N. He	ISE	Asst. Prof.	Large-scale Optimization
R. Nagi	ISE	Director of ISE	Large-scale Optimization
H. Hashemi	CS	Grad. Student	Machine Learning
R. Rutenbar	CS	Prof/Head of CS	Machine Learning
S. Koyejo	CS	Asst. Prof.	Machine Learning
J. Hockenmaier	CS	Assoc. Prof.	Natural Language Processing
M. Sammons	CS	Research Asst. Prof.	Natural Language Processing
L. Schwartz	Linguistics	Asst. Prof.	Natural Language Processing
L. Paquette	Education	Asst. Prof.	Social Network Analysis
Y. Hashash	CEE	Prof.	Soil Mechanics
K. McHenry	NCSA	Senior Research Scientist	Spatial Data Analysis
A. Das	ECE	Grad. Student	Speech Recognition
M. H. Johnson	ECE	Prof.	Speech Recognition
P. Smaragdis	CS	Asst. Prof.	Speech Recognition
W. Gropp	NCSA/CS	Director, NCSA/Prof. CS	Systems & Software
I. Gupta	CS	Assoc. Prof.	Systems & Software
D. Katz	NCSA	Asst. Director Sci. Software & Appl.	Systems & Software
V. Kindratenko	NCSA/ECE	Senior Research Scientist	Systems & Software
K. Nahrstedt	CS	Prof.	Systems & Software
R. Campbell	CS	Prof/Assoc. Dean IT	Systems & Software
S. Lazebnik	CS	Assoc. Prof.	Text mining
J. Han	CS	Prof.	Text Mining
W.M. Hwu	ECE	Prof.	Text Mining
C. Zhai	CS	Prof.	Text Mining

International collaborations

Cypress - joint PostDoc

RIKEN - FPGA development

Barcelona - scalable DL frameworks

System

P9 + NVIDIA Volta GPUs + HDR IB

Schedule

Year #	dates
1	August 2017 - August 2018	A single-node prototype based on the existing technology will be constructed first, including the supporting interconnect and storage infrastructure. This prototype will be used in the first year to investigate application requirements and study technology, such as NVM and FPGA choices, DRAM size, off-node bandwidth needs per GPU. Driven by user case studies, system-level software development for tightly integrating GPUs, NVM, and IB interconnect will begin. Limited number of domain scientists will be given access to the instrument to identify future instrument requirements and to gather data on the usage patterns. Development team will be hired and advisory board will be established.
2	August 2018 - August 2019	The majority of the instrument will be assembled, including GPUs and NVM; selection of an FPGA for future expansion will be completed. Storage system will be fully populated. Development of the software stack for seamless data movement across the storage hierarchies will continue, with first prototype supporting a limited set of DL frameworks. Resources allocation, user management, and system monitoring infrastructure will be fully deployed allowing multiple simultaneous users on the system. The system will be open to users capable of utilizing it via remote access over secure shell. Software development to enable easy user access via web portal and within environments such as Jupiter Notebook will begin. Widely used frameworks, such as Caffe and Tensor Flow, will be made available to the users, their optimization for the instrument will start as well. Initial training and outreach efforts will focus on training the users and developing documentation.
3	August 2019 - August 2020	System monitoring results will be used to optimize the system. A set of DL frameworks identified by the user community will be optimized to make use of the entire instrument, enabling DL at scale. Policies for instrument access and allocation will be developed and implemented with the help of the advisory board. FPGA-based hardware acceleration will be enabled for key computationally intensive tasks. Web portal for easy user access and Jupiter Notebook interface will be completed. Work will begin on other interfaces, such as R and Mathematica, driven by the user community needs. Software improvements and updates will continue throughout the year. System blueprints and performance results will be published in a paper. Access for education projects and courses will become available for a select set of users. By the end of year 3, instrument construction will be completed and the instrument will be transitioned for operation by the ISL at NCSA. Best practices for using the instrument and developing software for it will be developed and documented. The instrument will be open to the wide community of users, including off-campus researchers.

Child pages

NSF MRI: Development of an Instrument for Deep Learning Research