Scheduler benchmarks

Research results on real datasets, plus a live, reproducible replay of ASTRA PPO against classical baselines on the current cluster snapshot.

Research results (offline, on real datasets)

real trace

DRL-PPO scheduler

outperforms random placement

Google Cluster Trace 2011

49%

LSTM prewarming

fewer cold starts (N-RMSE 0.17)

Azure Functions 2019 trace

F1 0.82

Syscall IDS

LID-DS CVEs, beats STIDE and frequency

LID-DS 2021

30%

Carbon-aware

CO2 cut at a 24h deferral budget

UK Carbon Intensity API (live)

Scheduler explorer — compare all eight algorithms

Live scheduler explorer

Same workload, eight algorithms. They land on 0 different nodes.

CPU: 1.0 Mem: 1024Mi SandboxLanguage

Live simulator — 200 jobs, seed 42

Run a benchmark from the panel on the left.

Previous runs

0 logged

No runs yet. Each run you do is logged here.

ASTRA-IDE

How these numbers work (FAQ)

The Live simulator here runs on our project's own cluster model — it generates synthetic jobs and replays them against the current (simulated) cluster telemetry. It does not use an external dataset. The Research results cards at the top are the separate, offline evaluations on real datasets (Azure trace, Tetragon corpus, UK carbon API) that validate each breakthrough against its paper.