A Novel Reliability Index to Assess the Computational Resource Adequacy in Data Centers

Overview

The paper addresses how to right-size rack-level computing resources for the workload and for the reliability level the investor wants. The LOWP index combines the binomial outage probability of rack PSUs with the rack's load-duration curve. PSU availability is derived from IEEE Gold Book industrial rectifier data (MTTF 1,960,032 h, MTTR 16 h): pa = 0.999991836934606, qu = 8.163 × 10^-6. Rack power is built from server utilisation traces in the public Google cluster data set plus fan and PSU models.

The example rack has 10 blade servers rated 800 W each (400 W idle) and 40 local fans. PSU outages are modelled as k-out-of-n with a binomial distribution; in a 10-PSU rack (0.8 kW each), three simultaneous PSU failures leave 5.6 kW and the authors conclude three spare PSUs/servers give acceptable risk. With a 5.4 kW base load and 5.6 kW demand present 60% of the time, the example rack's LOWP is 3.9162 × 10^-14 %.

Losses in the power conditioning chain are also modelled as functions of utilisation. At rack level, average loss is 5.8% of rack rated power for the PDU, 2.8% for the UPS and 0.65% for the rack's 10 PSUs. At aggregate level the IPSS (UPS, PDU, PSU) consumes more than 10% of rated IT load (10.07 MW), and PDUs lose more energy in total than UPSs.

Limits: cable losses and redundant UPS-PDU paths are neglected; results rest on one published workload data set and generic PSU failure data; the index should be read together with the investor's target reliability and is not validated against measurements from an operating facility. The modular approach is said to extend to cooling loads.

Key points

1. PSU availability derived from IEEE Gold Book industrial rectifier data (MTTF 1,960,032 h, MTTR 16 h): pa = 0.999991836934606, qu = 8.1630653943 × 10^-6.
2. Rack PSU outages are modelled as k-out-of-n with a binomial distribution; in a 10-PSU rack (0.8 kW each), three simultaneous failures leave 5.6 kW and three spare PSUs/servers give acceptable risk.
3. With a 5.4 kW base load and 5.6 kW demand present 60% of the time, the example rack's LOWP is 3.9162 × 10^-14 %.
4. Average power loss at rack level is 5.8% of rack rated power for a PDU, 2.8% for a UPS and 0.65% for the 10 PSUs of a rack.
5. The IPSS (UPS, PDU, PSU) consumes more than 10% of rated IT load (10.07 MW) at aggregate level; the study neglects cable losses and redundant UPS-PDU paths.
6. The rack model assumes 10 blade servers rated 800 W each (400 W idle) with 40 local fans.

Using it in design

• Set the number of spare PSUs/servers in a rack with a LOWP-style calculation from target service reliability and the load duration curve, avoiding unnecessary overprovisioning.
• Model PDU, UPS and PSU losses as functions of utilisation in the power-chain energy budget, bearing in mind that PDU loss can exceed UPS loss.
• Use LOWP to identify which rack groups should run at high adequacy for latency-sensitive workloads.

Who should read it

Designers and operations planners sizing rack-level capacity, redundancy and power-chain losses.

Document contents

1. I. Introduction
2. II. State of the Art (A. Reliability-based computation resource adequacy, B. Power consumption modeling of IT loads and power supply components, C. Application of real data set in power analysis)
3. III. Methods and Procedures (A. Loss of workload probability (LOWP), B. Power consumption models of load sections, C. Google data set structure and processing)
4. IV. Results and Analysis (A. Power consumption models, B. Rack level computational resource)
5. V. Discussion and Contribution
6. VI. Conclusion and Recommendations

Related standards

• IEEE 493-2007

Read the document

Publication details

Authors

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

Publisher

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

Year

2021

Volume

9

Pages

54530-54541

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.1109/ACCESS.2021.3070915

Licence

CC BY 4.0 — share with attribution

Licence evidence

ieeexplore.ieee.org

Citation