Publications

My Google Scholar Page

My Dissertation

Providing Insight into the Performance of Distributed Applications Through Low-Level Metrics (2020)

• My advisor for this work was Dr. Jack Dongarra, and I was supervised by Dr. George Bosilca who was instrumental in guiding me through this work
• My primary contributions in this work are the introduction of Software-Based Performance Counters (SPCs) to the Open MPI runtime and using GPU metrics to create a new methodology for assessing GPU load for load balancing purposes
• The dissertation has five chapters: Introduction; Literature Review; MPI Performance Analysis and Tool Support Through Software-based Performance Counters; GPU Load Imbalance; and Conclusions

Conference Publications

Evaluating the Potential of Elastic Jobs in HPC Systems (PMBS Workshop at SC23)

• Authors: David Eberius, Md. Wasi-Ur Rahman, David Ozog
• I was the primary author for this work and conducted the bulk of the research, analysis, and writing
• The primary contribution of this work was a detailed analysis and evaluation of whether or not resizable (elastic) HPC jobs provide a significant benefit to several key system metrics: run time, wait time, turnaround time, and system utilization. In short, elastic jobs are capable of improving all of these metrics, even with relatively small percentages of elastic jobs on the system.

Understanding Strong Scaling on GPUs Using Empirical Performance Saturation Size (P3HPC Workshop at SC22)

• Authors: David Eberius, Philip Roth, and David M. Rogers
• I was the primary author for this work and conducted the bulk of the research, analysis, and writing
• The primary contribution of this work is in expanding the Roofline Model into what we call a Skyline Model by introducing problem size and GPU launch configurations as additional parameters. This creates a three-dimensional plane of peak performance on a system and can give additional insight into application performance on GPU-based systems.

Did the GPU Obfuscate the Load Imbalance in my MPI Simulation? (HiPar Workshop at SC21)

• Authors: David Eberius, David Boehme, Olga Pearce
• I was the primary author for this work and conducted the bulk of the research, analysis, and writing
• Olga Pearce came up with the idea for this research, and formulated the initial model for GPU load
• The primary contribution of this work is a new hierarchical metric for determining GPU load using low-level metrics from the nvprof profiler, which provides a much mor accurate assessment of GPU load than treating the GPU as one monolithic entity.

Multirate: A Flexibile MPI Benchmark for Fast Assessment of Multithreaded Communication Performance (ExaMPI Workshop at SC19)

• Authors: Thananon Patinyasakdikul, Xi Luo, David Eberius, George Bosilca
• My contribution to this work came in the form of my expertise in MPI performance analysis and the usage of the software-based performance counters in the Open MPI runtime
• I was the primary presenter for this research at SC19
• The primary contribution of this work is the Multirate benchmark, which is capable of stress testing both multithreaded and multiprocess MPI performance.

Give MPI Threading a Fair Chance: A Study of Multithreaded MPI Designs (CLUSTER 2019)

• Authors: Thananon Patinyasakdikul, David Eberius, George Bosilca, Nathan Hjelm
• My contribution to this work came in the form of my expertise in MPI performance analysis and the usage of the software-based performance counters in the Open MPI runtime
• The primary contribution of this work is in exposing some of the key shortcomings of multithreaded MPI and proposing a number of potential optimizations and workarounds that can drastically improve the performance of multithreaded MPI.

Using Software-based Performance Counters to Expose Low-level Open MPI Performance Information (EuroMPI 2017)

• David Eberius, Thananon Patinyasakdikul, George Bosilca
• I was the primary author for this work and conducted the bulk of the research, analysis, and writing
• The primary contribution of this work is the introduction of Software-based Performance Counters (SPCs) to the Open MPI runtime system. These SPCs allow for low-level instrumentation of the MPI implementation itself and provide insight into how the MPI implementation is operating, which would otherwise be opaque to the user.