wtf is philly-trace-scripts?

philly-trace-scripts is a collection of analysis scripts that help researchers understand how Microsoft managed a large network of computers equipped with GPUs (specialized chips for heavy-duty math, often used to train artificial intelligence). Microsoft publicly released a dataset, called the Philly trace, that captures the real-world activity of one of their AI training clusters from 2017. This project takes that raw, messy data and turns it into clean summaries and visual charts. The work is split across three Jupyter Notebooks, which are interactive documents that combine code, notes, and output. First, one script figures out which physical machines were actually active in the cluster by looking at GPU usage logs, producing a corrected machine list. Second, another script takes the job logs and that corrected machine list to build a timeline of how the cluster was used over time. Finally, a third script takes those timelines and machine lists and generates visual charts showing how the machines were distributed and how GPU utilization changed across the cluster. The audience for this project is researchers or data scientists studying how large-scale AI infrastructure operates in the real world. For example, if you are trying to understand how often AI training jobs get interrupted, how long they wait for resources, or how efficiently a massive cluster of GPUs is utilized, these scripts help you extract those insights from Microsoft's data. One notable detail is that the author found discrepancies between their own analysis and the official results provided by Microsoft. The project documents this conflict, pointing to an open discussion about the data's accuracy. This transparency makes the scripts valuable not just for generating charts, but for anyone who needs a trusted, independently verified understanding of how the original AI training cluster actually performed.