Skip to content
plexoosPublic

About

Efficient runtime execution for heterogeneous systems

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Teerex studies how to batch and dispatch event workloads across CPUs and a GPU. It combines a synthetic scheduling experiment, measured Simphony timelines, Bayesian phase-duration models, and offline policy replay.

The current question is how much CPU concurrency and work in flight are needed to feed the GPU while controlling queueing and completion latency. The study now uses the ratio-study-v3 production sample: 22,000 retained events across dRICH, pfRICH, balanced argon, and two GPU-heavy argon conditions. Each workload has a separate fitted model and validation-selected policy. Predicted asynchronous gains remain estimates to validate in an executor.

Start with the study notebook

Requires Python 3.11+ and uv.

uv sync
uv run marimo edit notebooks/simload.py

The main notebook tells the study's story and provides comparison plots, model checks, timelines, and interactive replay. Its bundled data snapshot works without the external study archive or a GPU. Running the real synthetic workload requires CUDA-enabled PyTorch and an NVIDIA GPU.

To rebuild the notebook's evidence from the combined production CSV:

PYTHONPATH=. uv run python scripts/refresh_ratio_study.py \
  ../teerex-study/ratio-study-v3/all-production-events.csv
PYTHONPATH=. uv run python scripts/refresh_ratio_study.py --verify

Documentation

The multithreaded Geant4 timeline notebook compares measured 1/2/4/8-worker runs across seven workloads and four seed pairs. Its bundled CSV snapshot preserves all 57,792 raw events, 56,000 retained events, run summaries, and timing caveats without requiring the external measurement archive or a GPU. The MT findings and scaling plots explain the measured gains, GPU-gate saturation, and the anomalous serial control. The browser version runs interactively without a local Python installation.

PYTHONPATH=. uv run --with plotly marimo edit notebooks/mt_timelines.py

To analyze a complete measured-run archive:

uv run study --root /path/to/teerex-study/drich-study-v2 --draws 64 \
  --workers 1 2 4 8 16 20 --inflight-limits 1 2 4 8 16 32 64

--root points to the archive containing runs/<run_id>/. --draws controls the posterior sampling budget; --workers and --inflight-limits define the simulated CPU worker counts and limits on unfinished events to compare. The two resource lists form a grid, so larger lists increase analysis time. Run uv run study --help for defaults and examples, or see the argument reference for input requirements and replay semantics.

make help lists development commands; make test runs the test suite.

About

Efficient runtime execution for heterogeneous systems

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Contributors

Languages