Phaedrus: Predicting Dynamic Application Behavior with Lightweight Generative Models and LLMs

2024-12-09

AI Summary

Modern applications exhibit highly input-sensitive behavior, rendering conventional single-execution profiling-based performance analysis ineffective. This paper proposes a compiler-coordinated deep learning framework that, for the first time, enables cross-input dynamic function-call sequence prediction without runtime profiling. Our approach comprises two core innovations: (1) an application-behavior synthesis paradigm integrating large language models with static analysis; and (2) a generative behavioral generalization method based on compressed whole-program paths (WPP), achieving up to 10⁷× profile size reduction. The framework achieves high-accuracy call prediction within 85–99% of critical execution time, enabling input-aware code optimization: binary size is reduced by 13.68% on average (up to 65%), while performance improves by 2.8% on average.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: Practical large-scale studies of user experience
• Systems and Infrastructure for Web, Mobile and WoT: Web performance, measurement, and characterization
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Application profiling is an indispensable technique for many software development tasks, such as code and memory layout optimizations, where optimization decisions are tailored to specific program profiles. Unfortunately, modern applications codebases exhibit highly variant behavior across different inputs, creating challenges for conventional profiling approaches that rely on a single representative execution instance. In this paper, we propose extbf{Phaedrus}, a new extit{compiler-assisted deep learning framework} designed to predict dynamic program behaviors across varied execution instances, specifically focusing on dynamic function call prediction.Such predicted call sequences are then used for producing optimized code pertinent to a given input. Traditional profile-guided optimization methods struggle with the input-dependent variability of modern applications, where profiling on different inputs yields divergent application behaviors. To address this, Phaedrus proposes two new approaches: extit{Application Behavior Synthesis}, a profile-less approach where Large Language Models (LLMs) directly infer dynamic functions based on source code &static compiler analysis, bypassing the need for traditional profiling, and extit{Application Profile Generalization}, which uses generative models trained on compressed and augmented extit{Whole Program Path} (WPP) based function profiles to predict application behavior under unseen inputs. Our experiments show that extit{Phaedrus} can achieve upto $10^7X$ reduction in WPP function profile sizes, can predict most frequently executed functions that cover upto 85-99% of the execution time, along with an average of 13.68% (upto 65%) reduction in application binary size, and an average of 2.8% performance improvement over the traditional profile-guided optimization.

Problem

• Predict dynamic program behaviors across varied execution instances
• Address input-dependent variability in modern applications profiling
• Optimize code using predicted function calls for specific inputs

Innovation

• Compiler-assisted deep learning predicts dynamic behaviors
• LLMs infer functions from source code statically
• Generative models generalize profiles for unseen inputs