xPUE: Extending Power Usage Effectiveness Metrics for Cloud Infrastructures

Overview

This paper, by Guillaume Fieni, Romain Rouvoy and Lionel Seinturier, was published in IEEE Transactions on Sustainable Computing in 2025 (pages 1-13, DOI 10.1109/TSUSC.2025.3549687, arXiv:2503.07124). The authors argue that facility PUE stops at the server power inlet and therefore hides losses inside servers (fans, power supplies, idle draw) and in the cloud software stack. They also stress that PUE counts only energy and not useful computational output.

The authors formalise server PUE (sPUE): server input energy divided by the energy of computing components such as CPU, DRAM, GPU and disks. A virtual PUE (vPUE) is defined for platform overhead. These multiply into a cloud PUE (cPUE) and, combined with facility PUE, a global gPUE: gPUE = cPUE x PUE and gCUE = cPUE x CUE, with gWUE = cPUE x WUE proposed by analogy. For measurement they used hardware power meters, IPMI, RAPL and a software power meter on Grid'5000 and OVHcloud servers.

The results show that idle servers gave an sPUE above 4. On a Dell R640 (Xeon Gold 5220) under rising CPU load the average sPUE was 3.1 and the lowest reached was 2.7. The best sPUE (1.4) came from an OVHcloud water-cooled server with a single AMD CPU; the authors attribute part of the liquid-cooling benefit to removing server fans, which PUE does not capture. At cluster scale (5 nodes), idle platforms gave sPUE above 20 (max 22.6 Kubernetes, 27.5 OpenStack).

In the software layer, OpenStack services took about 20% of energy (aggregated vPUE 1.25); combined with sPUE 2.7 the cPUE was 3.375, meaning 1 W in a hosted VM implied 3.375 W at cluster level. gPUE values across OVHcloud sites were all above 1.8, and gCUE varied by a factor of 26. As a limit, the measurements belong to specific hardware and platforms; the results should be interpreted with their scope stated, not generalised.

Key points

1. PUE can mechanically decrease by artificially increasing IT workload; it counts only energy and not useful computational output.
2. sPUE = Energy(IT) / Energy(hardware components doing the computation: CPU, GPU, memory, disk, controllers); gPUE = cPUE x PUE, gCUE = cPUE x CUE; gWUE = cPUE x WUE proposed by analogy.
3. Idle servers gave sPUE > 4; on a Dell R640 under rising CPU load the average sPUE was 3.1 and the lowest 2.7.
4. The best sPUE (1.4) came from an OVHcloud water-cooled server with a single AMD CPU; part of the benefit is attributed to removing server fans.
5. At cluster scale (5 nodes), idle platforms gave sPUE above 20 (max 22.6 Kubernetes, 27.5 OpenStack); the minimum was 2.9 and the stressed average 5 to 6.5 depending on platform.
6. A ToR switch drew a near-constant 115.3 W regardless of workload; the authors recommend counting network equipment in the IT numerator of sPUE.
7. OpenStack services took about 20% of energy (vPUE 1.25); with sPUE 2.7 the cPUE was 3.375; Kubernetes averaged vPUE 1.3 with an optimum of 1.1.
8. gPUE across OVHcloud sites was always above 1.8; gCUE spanned a factor of 26 (48.75 to 1,262 eqCO2 per unit of computation).

Using it in design

• Do not present facility PUE as the whole efficiency story: add an optional server-overhead factor (sPUE) to the energy model and show an 'effective PUE' = PUE x sPUE next to PUE.
• Warn when the IT load is modelled at low utilisation, because server-level overheads rise steeply when idle.
• When comparing air vs liquid cooling, move server fan power out of the IT load for liquid-cooled racks so the comparison is not biased by where fan energy is counted.
• Count network gear as IT load and report the scope used for each metric.

Who should read it

Designers interpreting PUE and efficiency metrics, cloud and colocation operators, and sustainability reporters.

Related standards

• ISO/IEC 30134-2

Publication details

Authors

Guillaume Fieni; Romain Rouvoy; Lionel Seinturier

Publisher

IEEE Transactions on Sustainable Computing (Institute of Electrical and Electronics Engineers (IEEE))

Year

2025

Document number

arXiv:2503.07124

Pages

1-13

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.1109/TSUSC.2025.3549687

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation