A Review of Data Centers Energy Consumption and Reliability Modeling

Overview

The study tackles two linked questions: how data-centre load sections are modelled for energy consumption, and how the reliability of the IT power conversion system (IPCS) is assessed. The method is a systematic literature review: energy models for the IT, internal power conditioning and cooling load sections are classified and their advantages and disadvantages set out by application, and reliability modelling approaches are then reviewed.

For energy modelling, the main conclusion is that the availability of model parameters and variables matters more than accuracy. On reliability, the deterministic Uptime Tier approach is contrasted with probabilistic methods (RBD, fault tree, Markov, Monte Carlo, Petri nets). The deterministic Tier calculation ignores grid outage probability, differing component failure rates and random failure modes, and assumes constant failure and repair rates. IEEE Std 493-2007 (Gold Book) defines failure as loss of power to a PDU (or UPS distribution panel), so a single branch breaker trip that de-energises one rack is not a data-centre failure; the failure state must therefore be defined before any study.

Five IPCS reliability indices are compiled from the literature: MTBF, MTTR, availability, severity and risk (financial loss caused by the failure). Cooling-system operational availability is expressed as AO = 1/(λsys/µsys + 1) and functional availability as Af = AO(1 − pus) + (1 − AO)·pt. Because Tier IV availability must reach five to six nines, very few failure events appear in a million stochastic samples, and Monte Carlo results depend strongly on failure-mode modelling and the chosen distributions.

Limits: as a review it has no new measurements; failure and repair data for PDUs, PSUs and cooling devices are not public, largely because operators keep them confidential, and only server failure and repair times have a public data set. Models therefore lean on generic data and assumptions, and the research gaps are listed as recommendations in the conclusion.

Key points

1. The deterministic Tier availability calculation ignores grid outage probability, differing component failure rates and random failure modes in the supply paths, and assumes constant failure and repair rates.
2. IEEE Std 493-2007 (Gold Book) defines failure as loss of power to a PDU (or UPS distribution panel); a single branch breaker trip that de-energises one rack is not a data-centre failure under that definition, so the failure state must be defined first.
3. Five IPCS reliability indices are compiled: MTBF, MTTR, availability, severity and risk (financial loss caused by the failure).
4. Cooling-system operational availability is AO = 1/(λsys/µsys + 1) and functional availability is Af = AO(1 − pus) + (1 − AO)·pt.
5. Tier IV availability must reach five to six nines, so very few failure events appear in a million stochastic samples; Monte Carlo results depend strongly on failure-mode modelling and chosen distributions.
6. A public data set exists for server failure and repair times, but failure and repair data for PDUs, PSUs and cooling devices is not public, largely because operators keep it confidential.
7. Conclusion: the availability of model parameters and variables matters more than accuracy for energy models.

Using it in design

• Define the failure state (loss of PDU power, loss of a rack, loss of service) explicitly before starting a reliability study and state it in the report.
• When showing Tier-level targets with a deterministic calculation, state that grid outages and failure-rate differences are not included; support it with a probabilistic method (Markov, Monte Carlo) and present failure-mode assumptions as inputs.
• Use the functional availability formula that combines cooling availability with room-temperature probabilities in cooling redundancy checks.
• Flag the source (manufacturer, operating record) and uncertainty for PDU, PSU and cooling failure data.

Who should read it

Researchers, design engineers and operations planners doing data-centre reliability and energy modelling.

Document contents

1. I. Introduction (A. Background, B. Research gaps, C. Objective and approach, D. Contributions and recommendations, E. Organization)
2. II. Related Review Articles
3. III. Review of Data Center Load Modeling (A. IT load models, B. Internal power conditioning system model, C. Cooling section models)
4. IV. Review of the Reliability Modeling of Data Centers (A. Tier classification, B. Factors to consider for the reliability, C. Reliability indices and metrics, D. Methodologies used for reliability modeling, E. Dependability of the data center load)
5. V. Limitations and Future Works
6. VI. Conclusion and Recommendations

Related standards

• IEEE 493-2007
• Uptime Institute Tier Standard: Topology

Read the document

Publication details

Authors

Kazi Main Uddin Ahmed; Math H. J. Bollen; Manuel Alvarez

Publisher

IEEE Access (Institute of Electrical and Electronics Engineers (IEEE))

Year

2021

Volume

9

Pages

152536-152563

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.1109/ACCESS.2021.3125092

Licence

CC BY 4.0 — share with attribution

Licence evidence

ieeexplore.ieee.org

Citation