Virtualization: Optimized Power, Cooling, and Management Maximizes Benefits

Overview

The paper lists four consequences for the infrastructure: localized high-density hot spots, a PUE that degrades once IT load falls relative to the plant, loads that move in time and place, and the option to reduce redundancy where software fault tolerance is strong. The starting point is server utilization. A non-virtualized server typically sits at 5-10% CPU use while a virtualized host may reach 50% or more, yet the power draw difference between those two utilization levels is only about 20%.

The energy example makes this concrete. At $0.11/kWh, 1000 servers drawing 250 W each cost about $240,000 a year. Consolidated 10:1 onto hosts drawing 600 W each, the cost falls to about $60,000, a 76% saving in server energy. The paper also mentions self-contained micro data centers, 1-100 kW of IT load, as one outcome of densification.

The core case study follows a 1 MW facility through three states. Before virtualization it runs at PUE 2.00 with 700 kW of IT load (70% loaded). Consolidation alone drops the load to 515 kW (52% loaded) and, because fixed losses stay, PUE rises to 2.25. Rightsizing power and cooling to 515 kW of capacity (100% loaded) brings PUE to 1.63. The electric bill falls 17% with consolidation alone and 40% when the UPS, PDUs and cooling are also rightsized and containment is applied. In the case, 90% of the savings come from optimizing the existing system and 10% from rightsizing; the UPS assumption is 89% efficiency at full load for a traditional unit against 96% for a high-efficiency one.

The paper also lists the operating hazards of severe underloading. Chillers can trip compressors on high head pressure, short-cycle, and incur hot-gas bypass cost; generators suffer wet stacking from unburned fuel, need jacket water heaters and extra fuel maintenance. In a highly virtualized, fault-tolerant IT environment it suggests designers may consider N+1 rather than 2N power and cooling. The evidence rests on a single 1 MW scenario, so the numbers illustrate the mechanism rather than predict a specific site.

Key points

1. Non-virtualized server CPU utilization is typically 5-10%; a virtualized host can reach 50% or higher, while the power draw difference is about 20%.
2. 1000 servers x 250 W at $0.11/kWh cost about $240,000/yr; 10:1 consolidation at 600 W per host cuts it to about $60,000, a 76% server energy saving.
3. 1 MW case: PUE 2.00 (700 kW, 70% loaded) before, 2.25 (515 kW, 52%) after consolidation only, 1.63 after rightsizing.
4. The electric bill drops 17% with consolidation alone and 40% when UPS, PDUs and cooling are rightsized and containment is added.
5. In the case, 90% of savings come from optimizing the existing system and 10% from rightsizing.
6. UPS assumption: traditional 89% at full load versus high-efficiency 96%.
7. Underloading hazards for chillers: high head pressure compressor shutdown, short-cycling, hot-gas bypass cost; for generators: wet stacking, jacket water heaters, extra fuel maintenance.
8. Micro data centers are self-contained enclosures of 1-100 kW IT load.
9. Strong software fault tolerance may allow N+1 instead of 2N.

Using it in design

• Model PUE as a load-dependent curve of fixed plus proportional losses and show it at day-1, partial and full IT load rather than as one number.
• Warn when day-1 IT load is below roughly 50% of installed power and cooling capacity and no modular or phased build plan exists; cite the fixed-loss penalty, compressor short-cycling and generator wet stacking.
• Record UPS efficiency at 25/50/75/100% load plus a modular flag, and a variable-speed fan flag on cooling units.
• Let users tag workloads as virtualized or fault-tolerant and present the N+1 versus 2N capex and efficiency trade-off instead of lowering redundancy silently.

Who should read it

Facility designers and energy managers deciding how to downsize infrastructure after virtualization or how to phase a new build.

Related standards

• ISO/IEC 30134-2

Publication details

Authors

Suzanne Niles; Patrick Donovan

Publisher

Schneider Electric

Year

2015

Document number

WP 118

Document language

English

Source type

Manufacturer

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation