Research

This project investigates advanced features of the AMD MI300A accelerator, including FP8 matrix cores, asynchronous execution engines, and structured sparsity. Through detailed performance characterization and benchmarking, the study evaluates how these hardware capabilities can accelerate AI and scientific computing workloads. The results provide insights that help researchers optimize applications and efficiently utilize next-generation HPC systems....

This project develops analytical models to predict application performance on modern GPU architectures such as NVIDIA Blackwell and AMD MI300A. By combining microbenchmarking with hardware characterization, the framework provides accurate performance estimates for scientific and AI workloads. The models help researchers optimize applications for next-generation high-performance computing systems. ...

This project presents a microbenchmarking study of NVIDIA's Blackwell GPU architecture, characterizing its memory hierarchy, compute throughput, and key architectural changes relative to prior generations. Using targeted microbenchmarks, we uncover low-level hardware behaviors that are not disclosed in vendor documentation. These findings give HPC developers and performance modelers a clearer picture of Blackwell's design for optimizing next-generation GPU workloads. ...

This project develops a suite of unit tests that validate and verify OpenACC compiler implementations against the official specification. Each test targets a specific directive or clause, checking that compiler behavior conforms to the standard across supported platforms. The testsuite helps compiler developers catch specification violations early and gives the OpenACC community shared benchmark for conformance....

This project addresses a critical gap: while LLMs have become powerful tools for code generation, verifying the correctness of LLM-generated code remains a challenge without comprehensive, autonomous solutions. LLM4VV explores using LLMs-as-judges to validate and verify compiler test suites, enabling more rigorous testing of parallel programming frameworks that have predominantly been a manual process and still continues to be....

This project, in collaboration with the FNLCR and NCI/NIH, produced UNNT (Utility for comparing Neural Networks and Tree-based models), an open-source framework that lets cancer-research scientists drop in their own RNA-Seq gene expression and drug-response tabular data and rigorously benchmark CNN and XGBoost models side by side on the same workflow. ...

In IEEE Proceedings of third Workshop on Accelerator Programming Using Directives (WACCPD), pp. 79-88, Salt Lake City, UT, November 2016....

Development of an accelerated version of the prediction of chemical shift of protein structures on GPUs using OpenACC on GPUs. This is the first directive-based version of the software that is available....

To appear in Proceedings of the Seventh International Workshop on Accelerators and Hybrid Exascale Systems (AsHES). IEEE Press, 2017....

This repository contains software developed using a portable, high-level framework using a popular MapReduce framework, Apache Spark, in conjunction with CUDA and OpenCL to take advantage of automatic data distribution and specialized hardware distributed across systems...

Undergraduate research project conducted within the Vertically Integrated Projects (VIP) program at the University of Delaware. Won 1st place at the VIP Mid-Atlantic poster competition, 2018....

In this code, we carefully utilize OpenACC, adirective-based programming model to accelerate the diffusion portion in Equation 1 inPhysiCell, a cross-platform agent-based biosimulationframework that has been adopted in cancer infectious diseases and othercomplex biological problem....

Work in Collaboration with Oak Ridge National Lab. Accelerating DOE’s Minisweep miniapp on supercomputers. Accepted, To be published in ACM Proceedings of The Platform for Advanced Scientific Computing (PASC) 2018....