Srihari Babu Godleti Publishes Research on LLM-Guided Optimization of Cloud Analytics

New research examines how large language models can help organizations balance performance, scalability and cost across cloud analytics platforms

SAN JOSE, Calif. — September 10, 2026 — Research by data and AI engineering professional Srihari Babu Godleti examining the use of large language models to optimize cloud analytics workloads has been published in the International Journal of Computational and Experimental Science and Engineering (IJCESEN).

The research article, “LLM-Guided Cross-Platform Optimization of Cloud Analytics Workloads,” was published in Volume 12, No. 2 of the journal on April 25, 2026. The article introduces an optimization framework called LLM-TradeOpt, designed to use workload characteristics, system configurations and historical execution information to develop adaptive strategies for cloud analytics environments.

The study addresses a practical challenge facing organizations operating large-scale analytics systems: improving processing performance while also controlling infrastructure costs and maintaining scalability.

Cloud analytics platforms can behave differently depending on workload characteristics and configuration choices. The research therefore evaluates LLM-TradeOpt across three different environments—Amazon EMR, Apache Spark running on Kubernetes and Snowflake—using workloads from the CloudSuite v4.0 benchmark suite.

According to the study’s reported results, LLM-TradeOpt achieved up to 18.7% lower latency, 22.4% higher throughput and 15.3% cost savings compared with the baseline approaches evaluated in the research.

“The objective was to explore whether large language models could reason across workload characteristics, platform behavior and cost rather than treating cloud optimization as a single-variable problem,” said Srihari Babu Godleti. “The results suggest that contextual, iterative optimization can provide a practical way to balance competing performance and economic objectives.”

The framework approaches optimization as a multi-objective problem involving latency, throughput and monetary cost. Rather than optimizing one metric independently, the research assigns configurable weights to those objectives and uses execution feedback to refine subsequent configuration recommendations.

The study also examined how the framework responded to different operational priorities. Under a performance-focused configuration, the research reported an average latency of 118.9 seconds and throughput of 531 queries per hour, while configurations placing greater emphasis on cost reduction produced progressively lower costs with corresponding changes in latency and throughput.

Beyond overall performance, the research evaluated execution stability and resource utilization. The reported results showed an 11.2% reduction in latency variance, while CPU utilization increased to 78% and memory utilization remained at 81% under the proposed framework.

The research further tested the approach across two different workload domains: Data Analytics and Data Serving. In the reported comparison, LLM-TradeOpt reduced latency by 16.9% in the Data Analytics domain and 20.4% in the Data Serving domain relative to the best-performing baseline in each domain. The study also reported cost reductions of 13.8% and 17.1%, respectively.

A key element of the research is its use of iterative reasoning. The framework generates candidate configurations, evaluates their predicted utility and constraint risks, and incorporates observed performance changes into subsequent optimization steps. The paper’s ablation study found that removing the iterative feedback component resulted in a 21.6% increase in latency and a 17.9% decrease in throughput relative to the full model.

Godleti’s publication builds on his broader work in data engineering and distributed computing. His earlier research includes “Taming Spark Data Skew with Practical Solutions,” published in the Journal of Computer Science and Technology Studies in June 2025. That paper examines repartitioning, key salting and broadcast joins as approaches to mitigating data-skew problems in Apache Spark. The publication is independently listed by the journal with DOI 10.32996/jcsts.2025.7.89.

His research on cloud data engineering has also included work examining the use of SonarQube and Snowflake for improving ETL processes, extending his focus across software quality, distributed data processing and cloud infrastructure.

Godleti’s professional work is focused on data and AI engineering, including enterprise data platforms, cloud technologies, distributed processing and emerging AI systems. His professional background includes engineering and technology roles at Roku, Amazon Web Services and Nike, among other organizations.

The newly published LLM-TradeOpt research reflects a growing area of interest in applying generative AI techniques not only to end-user applications but also to the infrastructure that supports modern analytics. As cloud environments become increasingly heterogeneous, the study proposes that AI-assisted reasoning could provide another mechanism for adapting system configurations to changing workload and business requirements.

About Srihari Babu Godleti

Srihari Babu Godleti is a data and AI engineering professional whose work focuses on cloud data platforms, distributed computing, analytics and artificial intelligence. His research interests include cloud optimization, large-scale data processing, Apache Spark, ETL systems and the application of large language models to data engineering problems.

Media Contact

Srihari Babu Godleti
Email: godleti.srihari@gmail.com
ORCID: 0000-0002-5247-7440
Research publication: LLM-Guided Cross-Platform Optimization of Cloud Analytics Workloads
DOI: 10.22399/ijcesen.5181

For more press release distribution and guest post services, visit TS Newswire.