Scheduler
tasqnetwork.io/benchmark/scheduler
A discrete-event simulation of mode R with hidden audits on 300 nodes across three GPU speed classes. 10% of nodes are adversarial and 5% are flaky with a 15% fault probability. Each run processes 20,000 units with r = 3, q = 2 and audit rate π = 0.1. Seat selection is either uniform at random or weighted by κ (reputation, price and latency).
12%Lower mean latency with κ at 50% load
0.1%Wrong acceptance in the last quarter of a run
67 to 87%Adversarial nodes paused per run
0%Honest nodes paused in any run
Mean latency by load
Randomκ-weighted
κ helps at moderate load. From 85% load upward every policy queues and κ is slightly slower than random in this model.
Source: sim_scheduler.csv, sim_scheduler.py
Wrong acceptance falls as reputation learns
First quarter of runLast quarter of run
κ-weighted selection. Audit catches pause adversarial nodes, so later units see fewer of them.
All runs
| Policy | Load | Mean latency | p95 latency | Utilisation | Wrong accept | Audit catches | Adversarial paused | Flaky paused | Honest paused |
|---|---|---|---|---|---|---|---|---|---|
| Random | 0.5 | 0.91 s | 2.72 s | 58% | 0.59% | 11 | 77% | 53% | 0% |
| κ-weighted | 0.5 | 0.80 s | 2.48 s | 49% | 0.56% | 10 | 67% | 27% | 0% |
| Random | 0.7 | 0.87 s | 2.65 s | 76% | 0.34% | 12 | 83% | 47% | 0% |
| κ-weighted | 0.7 | 0.82 s | 2.53 s | 71% | 0.68% | 13 | 87% | 20% | 0% |
| Random | 0.85 | 1.28 s | 3.11 s | 88% | 0.70% | 12 | 77% | 27% | 0% |
| κ-weighted | 0.85 | 1.42 s | 3.35 s | 88% | 0.85% | 11 | 73% | 40% | 0% |
| Random | 0.95 | 6.11 s | 12.33 s | 90% | 0.69% | 10 | 67% | 53% | 0% |
| κ-weighted | 0.95 | 6.34 s | 12.77 s | 89% | 0.65% | 12 | 77% | 60% | 0% |
Source: sim_scheduler.csv