arXiv subjects

Srinivasan Ramesh

Publications and source records attributed to Srinivasan Ramesh.

4 recordsLinked to original sources

arXiv2026

KEET: Explaining Performance of GPU Kernels Using LLM Agents

Performance profiles of GPU kernels generated by tools such as Nsight Compute are rich in detail but are often challenging to interpret. To achieve the best performance possible on a given GPU architecture, kernel developers need to spend significant time analyzing and comparing profiles in the tool's graphical interface to identify and understand kernel performance bottlenecks. Large Language Models (LLMs) have shown promise in understanding complex data and generating natural language explanations. In this paper, we propose the Kernel Execution Explanation Toolkit (KEET), an LLM-based agentic framework for interpreting Nsight Compute profiles to generate useful and data-grounded natural language explanations of performance issues in GPU kernels, and suggestions for optimizations. We evaluate \toolname using several CUDA kernels of varying complexity on NVIDIA H100 GPUs. We find that the generated explanations, when provided as context, improve the quality of LLM code optimization and multiple-choice question answering in downstream tasks. We further demonstrate that the tool can be used to interpret performance data from large sets of profiles to improve the quality of optimization suggestions.

cs.PF↗

arXiv2022

HPC Storage Service Autotuning Using Variational-Autoencoder-Guided Asynchronous Bayesian Optimization

Distributed data storage services tailored to specific applications have grown popular in the high-performance computing (HPC) community as a way to address I/O and storage challenges. These services offer a variety of specific interfaces, semantics, and data representations. They also expose many tuning parameters, making it difficult for their users to find the best configuration for a given workload and platform. To address this issue, we develop a novel variational-autoencoder-guided asynchronous Bayesian optimization method to tune HPC storage service parameters. Our approach uses transfer learning to leverage prior tuning results and use a dynamically updated surrogate model to explore the large parameter search space in a systematic way. We implement our approach within the DeepHyper open-source framework, and apply it to the autotuning of a high-energy physics workflow on Argonne's Theta supercomputer. We show that our transfer-learning approach enables a more than $40\times$ search speedup over random search, compared with a $2.5\times$ to $10\times$ speedup when not using transfer learning. Additionally, we show that our approach is on par with state-of-the-art autotuning frameworks in speed and outperforms them in resource utilization and parallelization capabilities.

cs.DC↗

arXiv2022

The Ghost of Performance Reproducibility Past

The importance of ensemble computing is well established. However, executing ensembles at scale introduces interesting performance fluctuations that have not been well investigated. In this paper, we trace our experience uncovering performance fluctuations of ensemble applications (primarily constituting a workflow of GROMACS tasks), and unsuccessful attempts, so far, at trying to discern the underlying cause(s) of performance fluctuations. Is the failure to discern the causative or contributing factors a failure of capability? Or imagination? Do the fluctuations have their genesis in some inscrutable aspect of the system or software? Does it warrant a fundamental reassessment and rethinking of how we assume and conceptualize performance reproducibility? Answers to these questions are not straightforward, nor are they immediate or obvious. We conclude with a discussion about the performance of ensemble applications and ruminate over the implications for how we define and measure application performance.

cs.DC↗

arXiv2017

Deep and Shallow convections in Atmosphere Models on Intel Xeon Phi Coprocessor Systems

Deep and shallow convection calculations occupy significant times in atmosphere models. These calculations also present significant load imbalances due to varying cloud covers over different regions of the grid. In this work, we accelerate these calculations on Intel{\textregistered} Xeon Phi{\texttrademark} Coprocessor Systems. By employing dynamic scheduling in OpenMP, we demonstrate large reductions in load imbalance and about 10% increase in speedups. By careful categorization of data as private, firstprivate and shared, we minimize data copying overheads for the coprocessors. We identify regions of false sharing among threads and eliminate them by loop rearrangements. We also employ proportional partitioning of independent column computations across both the CPU and coprocessor cores based on the performance ratio of the computations on the heterogeneous resources. These techniques along with various vectorization strategies resulted in about 30% improvement in convection calculations.

cs.DC↗