CPU-Only Benchmark

This page reports results from the CPU-only cluster benchmark experiment:

• experiments/01-cpu-only-benchmark/

Scope

The benchmark compares three baselines on a pure CPU cluster:

• A: simulator only (Joulie-free)
• B: Joulie with static partition policy
• C: Joulie with queue-aware policy

It evaluates energy and throughput under real Kubernetes scheduling with KWOK nodes and simulated power control.

Experimental setup

Cluster and nodes

• kind control-plane + worker (real control plane)
• 8 managed KWOK nodes - CPU only, no GPUs
• Workload pods target KWOK nodes via selector + toleration

Node inventory

Total: 8 nodes, 2304 CPU cores, 0 GPUs.

Hardware models in simulator

CPU power per node:

P(u, f) = IdleW + (PeakW - IdleW) * u^AlphaUtil * f^BetaFreq

Full power-model details: Power Simulator

Run configuration

• Seeds: 3
• Jobs: 300
• Mean inter-arrival: 0.20 s
• Time scale: 60x
• Timeout: 1800 s
• Perf ratio: 20%, GPU ratio: 0%
• Workload types: cpu_preprocess, cpu_analytics
• Policy caps: CPU eco at 65% of peak

Algorithms used

Controller policies

• static_partition:
  • hpCount = round(N * 0.30) -> 2 performance nodes, 6 eco nodes
• queue_aware_v1:
  • baseCount = round(N * 0.30), dynamic from live perf-pod count
  • hpCount = clamp(max(baseCount, queueNeed), 2, 15, N)
• Downgrade guard: performance -> eco deferred while performance-sensitive pods still run on node

Results summary

Per-seed results

Baseline means (3 seeds, all completed)

Relative to A:

• B: energy -31.8%, throughput -1.1% (negligible)
• C: energy -31.9%, throughput -0.8% (negligible)

Plot commentary

Runtime distribution

• All three baselines complete in nearly identical wall-time windows.
• Run-to-run seed jitter is larger than any inter-baseline difference.

Energy vs makespan

• B and C are consistently lower-energy than A with near-identical makespan across all 3 seeds.
• Both Joulie baselines cluster tightly together.

Baseline means

• Energy is the main differentiator; throughput and wall-time bars are indistinguishable.

Completion summary

• All 3 seeds completed for all baselines; no timeouts or gang-scheduling issues.

Interpretation

Joulie reduces energy by ~32% without throughput penalty on a CPU-only cluster because:

1. The cluster is over-provisioned (2304 cores, lightweight jobs) - eco nodes have spare CPU cores to compensate for throttled frequency.
2. CPU sensitivityCPU for cpu_preprocess/cpu_analytics is moderate (0.7-0.9): a 35% frequency reduction causes 25-32% per-job slowdown, but job completion time stays flat because the scheduler redistributes load.
3. Eco nodes draw significantly less power for the same simulated duration -> energy falls without extending makespan.
4. The aggressive 65% eco cap maximizes power savings on eco nodes compared to milder caps.

Best-fit use case

The strongest observed benefit is:

• energy reduction (-31.8% static, -31.9% queue-aware) with negligible throughput penalty in CPU-only mixed workload clusters.

Both policies perform equivalently on CPU-only clusters. static_partition is simpler to configure; queue_aware_v1 becomes more valuable when the performance-sensitive fraction is larger or more bursty.

Implementation details and scripts

• Experiment folder
• Main scripts:
  • 05_sweep.py
  • 06_collect.py
  • 07_plot.py
• Full report