Specpower Eval

specpower-eval

Towards Energy-Proportional Computing Using Subsystem-Level Power Management — Subramaniam and Feng (2015) (arXiv:1501.02724, 2015)

What this evaluates

Evaluates the energy proportionality and power efficiency of enterprise server subsystems under varying JVM-based web service workloads. It measures how power consumption scales with workload intensity and identifies non-proportional power draw in uncore components.

Datasets

• SPECpower_ssj2008 — total ?; splits: test (-1)

Metrics

• watts (primary) — range: watts
  • Measured directly using a Yokogawa WT210 power meter at the system level. Reported at idle and at 100% load-level (target 1,120,000 ssj_ops).

Input / output format

Input: System Under Test (Intel Xeon E5-2665, 256 GB RAM, Linux 3.2.0) running 8 JVMs with 4 threads each, pinned to adjacent physical cores, with HugeTLB enabled.

Output: Power consumption readings in watts, throughput in ssj_ops, and response time compliance metrics.

Scoring recipe

power_idle = measure_power()
power_peak = measure_power_at_load(100)
throughput = get_throughput() # ssj_ops
energy_prop = power_idle / power_peak
return power_idle, power_peak, throughput, energy_prop

Common pitfalls

• Uncore subsystem power remains constant regardless of workload load, skewing energy proportionality calculations if only core power is considered.
• Idle power consumption is significant (36–55% of peak), contradicting assumptions of near-zero idle power in traditional servers.
• Disk I/O must be eliminated (using RAMFS) to isolate processor and memory power management effects.

Evidence (verbatim from paper)

On an average, the SUT consumes 120 watts at idle and 330 watts at 100% load-level of SPECpower.

Citation

@misc{subramaniam2015towards,
  title={Towards Energy-Proportional Computing Using Subsystem-Level Power Management},
  author={Subramaniam and Feng (2015)},
  year={2015},
  note={arXiv:1501.02724}
}

• arXiv: 1501.02724