Project Info
Expanding GPU Memory with DRAGON-DIRECT
Mehmet Belviranli
belviranli@mines.edu
Project Goals and Description:
GPU Memory is often the limiting factor in general-purpose GPU compute problems. Our group is focused on expanding GPU memory space to use NVMe-backed buffers using DRAGON-DIRECT, specifically in data science and machine learning applications.
In this project we aim to integrate several GPU applications that require large memory so that they can benefit from DRAGON-DIRECT. The targeted applications include graph neural networks and CUDF based applications.
More Information:
Grand Challenge: Engineer the tools of scientific discovery.
N/A
Primary Contacts:
Mehmet Belviranli, belviranli@mines.edu | Benjamin Wagley, bwagley@mines.edu
Student Preparation
Qualifications
- A strong understanding of computing systems and Linux.
- Comfort with C/C++, CUDA, and Python in a headless Linux environment.
- Familiarity with Linux device drivers (optional)
- Familiarity with dataflow/ML and other data science paradigms (optional).
TIME COMMITMENT (HRS/WK)
10
SKILLS/TECHNIQUES GAINED
Students will learn about data science and machine learning system design for high performance computing. This may be in the form of improving driver-level system integration for stability and performance, or by developing novel integrations of DRAGON-DIRECT into existing data-science and ML libraries that build on a CUDA GPU stack.
MENTORING PLAN
There will be weekly meetings with the faculty advisor. There will also be a second weekly meeting with the PhD student managing the project as well as other students in the group.
PREFERRED STUDENT STATUS
Freshman
Sophomore
Junior
Senior