BranchBench – ROOT I/O & Compute Benchmark

A performance study comparing:

• Manual TTree event loop
• ROOT RDataFrame execution

Motivation

High Energy Physics workflows depend heavily on ROOT-based I/O. This project studies:

• Compute-bound scaling
• Framework abstraction overhead
• Manual vs RDataFrame performance behavior

Workload

• 1 scalar branch
• 1 vector branch (size 20)
• Gaussian-distributed values
• Heavy per-element trigonometric compute
• Deterministic random seed (42)

Build

mkdir build cd build cmake .. make

Run

./io_benchmark 500000 ./io_benchmark 5000000 ./io_benchmark 20000000

System Used for Benchmark

• Ubuntu 24.04 (WSL)
• g++ 13.3
• ROOT 6.x
• Single-thread execution

Scaling Results (Heavy Compute)

Events	Manual (ev/s)	RDataFrame (ev/s)
500,000	98k	76k
5,000,000	102k	95k
20,000,000	100k	96k

Observations

• Throughput stabilizes at ~100k events/s across large datasets.
• This indicates compute-bound behavior (linear time scaling with N).
• For small event counts (e.g. 500k), RDataFrame shows noticeable startup overhead.
• As dataset size increases, this fixed setup cost is amortized.
• For large datasets (20M events), RDataFrame overhead reduces to ~4–5%.
• In the steady-state regime, abstraction overhead becomes minor compared to per-event compute cost. Note: Relative differences are larger at small event counts due to framework initialization costs that are amortized at scale.