USCMS Researcher: Mason Proffitt



Postdoc dates: Jun 2026 - May 2027

Home Institution: Texas Tech University


Project: Streamlining inference with SONIC

This project aims to develop, test, and improve SONIC (Services for Optimized Network Inference on Coprocessors). SONIC provides a way to offload and batch processing of algorithms that can be processed much faster and more efficiently on heterogeneous resources beyond general-purpose CPUs (e.g., GPUs). This work includes facilitating the use of new models, optimization and performance benchmarking, and new server builds on AlmaLinux.

More information: My project proposal

Mentors:

  • Yongbin Feng (Texas Tech University)

Presentations
Current Status


2026 Q3

  • Isolated reconstruction step in MLPF performance tests to investigate discrepancy in producer execution time between cases with and without SONIC
    • Determined that difference was due to how times are reported for SONIC’s asynchronous mode (difference disappears in synchronous mode)
  • Prepared the GitHub branches needed for implementing the MLPF SONIC producer * Created pull request in FastML fork of CMSSW for review
    • Cleaned up code and factored out some common functionality between standard MLPF producer and SONIC version


2026 Q2

  • Followed tutorial from Yao to get acquainted with SONIC and CMSSW * Tested new MLPF model with perf_analyzer and in CMSSW
    • Testing with perf_analyzer suggested that performance could be modestly improved with concurrency greater than 1 and with multiple model instances
    • CMSSW timing tests showed discrepancy when running with SONIC versus without, to be investigated


Contact me: