Date: Friday, October 16th, 2026

Time: 1:00 - 3:00 PM ET

Location: Klaus 2126

Virtual Link: Teams (https://teams.microsoft.com/meet/293867738134584?p=GI8Tcf7EIQ4ih1fkL7

                                      Meeting ID: 293 867 738 134 584 

                                      Passcode: CN9DD3Zy)

Bodhisatwa Chatterjee
Ph.D. Student

School of Computer Science

College of Computing

Georgia Institute of Technology

Committee:

Dr. Santosh Pande (Advisor) – School of Computer Science, Georgia Institute of Technology

Dr. Qirun Zhang – School of Computer Science, Georgia Institute of Technology

Dr. Vijay Ganesh – School of Computer Science, Georgia Institute of Technology

Dr. Rajeev Gupta – Department of Computer Science & Engineering, University of California, Riverside

Dr. Mark Stephenson – NVIDIA Research, NVIDIA

Abstract

Accurate knowledge of program execution behavior is critical for both compiler optimizations and system-level resource management. However, modern workloads exhibit complex, input-dependent behavior with irregular control flow and dynamic phase changes, making static reasoning inadequate. While dynamic optimization techniques offer an alternative, their high analysis overhead often negates potential performance gains. In this work, we propose lightweight predictive techniques that leverage runtime values, program analysis, machine learning models, and LLM reasoning to predict dynamic program behavior at multiple levels of granularity. Although such predictions do not provide strict guarantees, we demonstrate empirically that they enable a range of “soft” optimizations that yield significant performance improvements.

In our first work, we showcase Compiler-Guided Cache Apportioning System (Com-CAS) for effectively apportioning the shared Last-Level Cache (LLC), through the use of runtime and compile-time cooperative framework. Com-CAS improves weighted throughput by 15% on average, outperforming the state-of-the-art schemes by 20% on average. In our second work, we extend this approach to estimate trip count for irregular loops, and present Beacons Framework, an end-to-end system that estimates dynamic loop characteristics and leverages them for proactive workload scheduling. Beacons achieves approximately 85--95% prediction accuracy and improves throughput by up to 2.62x over existing schedulers.

In our third work, we introduce Phaedrus, a framework for profile-driven software optimization via predictive modeling of dynamic application behavior. Phaedrus combines application profile generalization and application profile synthesis, integrating generative deep learning models and LLMs with static compiler analysis to predict function calls and control-flow behavior. This enables profile-guided optimization without exhaustive profiling across application inputs and demonstrates performance gains over traditional offline profiling approaches. Phaedrus improves performance by 6% on average and reduces binary size by 13.68%. 

Finally, we present Cassandra, a compiler-LLM co-reasoning framework that selectively predicts fine-grained, input-specific program behavior within computationally important regions identified through coarse-grained analysis. Cassandra first uses LLM reasoning to identify hot function chains, then applies compiler analyses and loop categorization to predict loop trip counts, invocation frequencies, and other intra-function behavior required for profile synthesis. Cassandra PGO produces leaner code than Traditional PGO and achieves aggregate runtime improvements of 1.32-1.44% over Traditional PGO.