Hailong Jiang

Hailong Jiang

he/him

Tenure Track Assistant Professor in Computer Science at Youngstown State University. Research focuses on HPC systems, compiler technologies, and the intersection of program analysis and large language models (LLMs).

HAPPA: HPC Application Resilience Analysis Platform

Published:

Project Overview

HAPPA (HPC Application Resilience Analysis Platform) is my flagship research project that represents a breakthrough in combining Large Language Models with high-performance computing resilience analysis. This work was presented at the prestigious SC’24 conference as part of the Doctoral Showcase and has been published in the SRDS’24 conference proceedings.

Key Innovations

LLM Integration for Code Understanding

  • Long Sequence Processing: Advanced LLM integration to understand and analyze long code sequences
  • Code Representation: Innovative techniques for representing HPC applications in a format suitable for LLM analysis
  • Semantic Analysis: Deep understanding of code semantics beyond traditional static analysis

Superior Predictive Performance

  • DARE Dataset: Utilized the comprehensive DARE (Dataset for REsilience analysis using Fault Injection) dataset for training and evaluation
  • MSE Achievement: Achieved mean squared error (MSE) of 0.078 in Silent Data Corruption (SDC) prediction
  • Benchmark Comparison: Significantly outperformed the existing PARIS model (MSE: 0.1172)
  • Accuracy Improvement: 30% improvement in resilience prediction accuracy

Technical Architecture

The platform consists of several sophisticated components:

  • LLVM Integration: Custom passes for soft error injection and analysis
  • Neural Network Models: BERT-based code understanding for resilience prediction
  • Real-time Monitoring: Continuous system health assessment during execution
  • Scalable Design: Support for large-scale HPC applications and distributed systems

Research Contributions

1. HPC Loop Resilience Analysis

  • 13 Dwarfs of Parallelism: Comprehensive analysis of computational patterns
  • SDC Rate Quantification: Quantified Silent Data Corruption rates for each pattern
  • Prompt Engineering: Advanced LLM prompting techniques for loop semantics identification
  • Error-Prone Loop Detection: Identified which loops are more susceptible to errors

2. IR Code Analysis with LLMs

  • Model Evaluation: Comprehensive testing of GPT-4o, GPT-3.5, and CodeLlama
  • Low-Level Analysis: Decompiling IR code, generating Control Flow Graphs (CFGs)
  • Code Execution Simulation: Advanced simulation capabilities for IR code analysis
  • Program Analysis Applications: Insights into LLM effectiveness for compiler-level tasks

3. Novel Code Representation Module

  • Code Chunking: Fixed-size segment division for long code sequences
  • Embedding Aggregation: Three aggregation methods explored:
    • MeanPooling technique
    • MaxPooling technique
    • LSTM-based aggregation (HAppA-LSTM)
  • Context Preservation: Maintains code context information across segments

4. KeyBERT Integration

  • Keyword Extraction: Automatic identification of source code key words
  • Importance Analysis: Comprehensive analysis of code patterns contributing to error rates
  • Pattern Recognition: Enhanced understanding of resilience factors

Impact and Recognition

  • Conference Publication: Published in SRDS’24 - International Symposium on Reliable Distributed Systems
  • Conference Presentation: Featured at SC’24 Doctoral Showcase in Atlanta, GA
  • Research Recognition: Significant contribution to AI-driven HPC resilience
  • Industry Relevance: Addresses critical challenges in system reliability and fault tolerance
  • Academic Contribution: Advances the intersection of artificial intelligence and high-performance computing

Future Applications

This platform serves as a foundation for:

  • Intelligent Fault Tolerance: AI-driven system reliability enhancement
  • Compiler Optimization: LLM-assisted code analysis and optimization
  • HPC System Design: Data-driven approaches to resilient system architecture
  • Research Collaboration: Framework for future HPC resilience research

Technical Details

  • Programming Languages: C++, Python, LLVM IR
  • ML Frameworks: PyTorch, Transformers, BERT, LSTM
  • Tools: LLVM/Clang, CUDA, OpenMP, KeyBERT
  • Platform: Linux-based HPC environments
  • Performance: Real-time analysis with minimal overhead

Publication Details

This work represents a significant milestone in my PhD research and demonstrates the potential of combining cutting-edge AI technologies with traditional HPC system analysis. The publication in SRDS’24 validates the technical contributions and establishes HAPPA as a benchmark for predictive accuracy in HPC resilience analysis.