• Final Project for Operating Systems Lab
• Instructor: Dr. Nabavi
• University: Shahid Beheshti University (SBU)

This project focuses on improving Readahead mechanisms in operating systems using data-driven modeling and machine learning.

Readahead is a technique used by the OS to prefetch pages into the page cache before a process actually requests them, in order to reduce I/O latency.
While this improves performance for sequential access patterns, static or poorly tuned readahead parameters can also cause cache pollution and inefficient I/O when the workload behavior changes.

The goal of this project is to:

• Analyze different I/O workloads and their effect on readahead performance
• Extract and learn workload-specific features from system traces
• Train models to dynamically determine optimal readahead sizes
• Implement and integrate an adaptive readahead system call into a teaching OS (e.g. Pintos)

• Collect low-level kernel trace data using LTTng while running RocksDB benchmarks.

• Tracepoints such as:

  • add_to_page_cache
  • writeback_dirty_page
  • Other page-cache-related functions

• Run at least 10 benchmarks, ensuring steady-state performance.

From collected traces, compute per-interval or per-transaction metrics such as:

• Average and variance of page faults per second
• Number of read/write transactions
• Differences between consecutive page offsets
• Moving averages and standard deviations

Use the trace information (e.g., struct page, inode, etc.) to design hooks for data collection.

• Normalize data using Z-Score normalization
• Reduce dimensionality or visualize using t-SNE
• Label workloads as one of four classes:
  • readseq
  • readreverse
  • readrandom
  • writerandom

• Train a neural network classifier to distinguish between workloads based on extracted features.
• Use:
  • 10-fold cross-validation
  • Cross-Entropy loss
  • Hyperparameter tuning for optimal accuracy
• Evaluate the model using Accuracy and confusion matrix.

• Train a Decision Tree model for interpretability.
• Visualize and compare its structure and classification performance with the neural network.

• Add a new system call to Pintos, e.g.:

  sys_readahead_tune(...)