Energy computation for ICT infrastructure
Open in the Padhaiko app6.1 Energy auditing for data centers and ICT infrastructure
An ICT energy audit systematically measures where electrical energy enters a data center or enterprise network, how much reaches useful computation, and where losses occur (cooling, UPS conversion, lighting, and idle servers).
Audit Methodology
- Preliminary Walk-Through Audit: Inventory server racks, UPS units, CRAC/CRAH cooling units, and utility electricity bills to spot obvious waste (blocking airflow, zombie/idle servers, uncontained aisles).
- Detailed Instrumented Audit: Measure active power (), power factor (), harmonic distortion (), rack inlet/outlet temperatures, and airflow rates across peak and off-peak intervals using power quality analyzers and DCIM sensors.
- Analysis & Benchmarking: Compute standardized efficiency metrics and rank energy-conservation measures (ECMs) by payback period and carbon reduction.
Key Data Center Energy Efficiency Metrics
PUE, DCiE, CUE, and WUE
- Power Usage Effectiveness (): Defined by The Green Grid as the ratio of total facility energy to IT equipment energy:
Ideal ; legacy data centers often run at , while modern hyperscale facilities achieve .
- Data Center Infrastructure Efficiency (): Reciprocal of expressed as a percentage:
- Carbon Usage Effectiveness (): Total emissions caused by the data center divided by IT energy ().
- Water Usage Effectiveness (): Annual site water consumption for evaporative cooling divided by IT energy ().
6.2 Thermal management and cooling requirements
Nearly of electrical power consumed by servers, switches, and indoor UPS losses turns into sensible heat inside the IT room. Failing to remove this heat causes thermal throttling, hardware failure, and unplanned shutdowns.
Cooling Load Calculation
Total indoor heat rejection load (in ) equals IT power plus indoor electrical losses, lighting, and occupant/envelope heat gain. Standard HVAC units express cooling capacity in Tons of Refrigeration () or :
For liquid or chilled-water loops, heat removal capacity depends on mass flow rate and coolant temperature rise :
Thermal Guidelines and Cooling Technologies
- ASHRAE TC 9.9 Guidelines: Recommended server inlet temperature envelope is (allowing higher cold-aisle setpoints so chillers work less and free cooling hours increase).
- Hot-Aisle / Cold-Aisle Layout & Containment: Racks face front-to-front (cold intake aisle) and back-to-back (hot exhaust aisle); physical containment barriers prevent hot exhaust air from recirculating into cold server intakes.
- CRAC vs. CRAH: CRAC (Computer Room Air Conditioner) uses direct-expansion refrigerant compressors; CRAH (Air Handler) circulates air over chilled-water coils fed by a high-efficiency central chiller.
- Economizers (Free Cooling): Air-side or water-side economizers bypass mechanical compressors whenever outside ambient air/water is cool enough (very effective in cool hill/mountain climates).
- Direct-to-Chip & Immersion Liquid Cooling: Required for high-density AI/GPU racks (), using water or dielectric fluids with far higher heat capacity than air.
6.3 ICT energy system design and power distribution
Mission-critical ICT facilities require clean, continuous, fault-tolerant power from utility transformers down to dual-corded server power supply units (PSUs).
Uninterruptible Power Supply (UPS) Topologies
| Topology | Normal Power Path | Transfer Time | Efficiency | Ideal Application |
|---|---|---|---|---|
| Offline / Standby | Utility feeds load directly; inverter starts only on outage | Desktop PCs, non-critical edge peripherals | ||
| Line-Interactive | Autotransformer regulates sags/surges (AVR); inverter stays synchronized | Small office switches, branch server closets | ||
| Online Double-Conversion | continuously | (zero interruption) | Data centers, telecom core, critical medical/network gear |
Power Distribution & Redundancy Tiers
Power flows through Main Switchgear + Automatic Transfer Switch (ATS) Backup Diesel/Fuel-Cell Generator UPS + Battery Bank Power Distribution Units (PDUs) Rack PDUs (rPDUs).
ICT Power Redundancy Architectures
- (No Redundancy): Exactly modules meet full load; any single failure causes downtime.
- : One spare module added alongside required modules; tolerates one component failure or maintenance window.
- (Fully Redundant A+B Feeds): Two completely independent power paths (Utility/UPS/PDU A and B) feeding dual-PSU servers; each path carries load normally so a complete path outage causes zero disruption.
- : Dual independent paths, each internally redundant (Tier IV fault-tolerant data centers).
Modern hyperscale facilities also adopt high-voltage DC distribution (e.g., in telecom or in data centers) to eliminate redundant conversion stages.
6.4 Energy aware algorithm design, energy complexity, green computing
Hardware efficiency alone cannot curb ICT energy growth; software and algorithms control how often processors, memory, disks, and radios become active.
Dynamic Voltage and Frequency Scaling (DVFS) and Power States
CMOS processor power has two components:
where is switching activity, is switched capacitance, is supply voltage, and is clock frequency. Because scales roughly with , reducing voltage and frequency yields an approximately cubic reduction in dynamic power ().
- DVFS vs. Race-to-Halt: When static leakage power () is small and the task has deadline slack, slowing and saves energy. When is high, running fast at peak efficiency and entering deep ACPI C-states (sleep states) (race-to-halt) can use less total energy.
Energy Complexity & Energy-Aware Algorithm Design
Traditional complexity measures time and space ; energy complexity models total joules , accounting for memory hierarchy costs:
- Minimize Expensive Data Movement: Fetching word from DRAM consumes roughly more energy than an on-chip ALU addition, and network/disk I/O costs orders of magnitude more. Cache-oblivious / cache-blocked algorithms, data compression, and in-memory batching sharply lower joules per operation.
- Avoid Busy-Waiting (Polling): Replace tight
while(!ready)spin loops with interrupt-driven event loops and OS sleep states. - Approximate & Mixed-Precision Computing: Using
FP16/INT8quantization in ML inference reduces memory bandwidth and MAC energy by with negligible accuracy loss.
Green Computing Lifecycle
Green computing spans the full ICT lifecycle: (1) Green Design (energy-proportional servers, Energy Star hardware), (2) Green Use (virtualization, container consolidation, dark-mode OLED UIs, cloud workload shifting to renewable-rich hours/regions), (3) Green Manufacturing (lead-free RoHS compliance, recycled aluminum/plastics), and (4) Green Disposal (certified e-waste recycling and component refurbishment).
Worked Examples
6.1
An enterprise data center operates with an IT server/network load , cooling plant power , indoor UPS/PDU conversion losses , and indoor lighting/auxiliary load . Calculate:
- Total facility power (), , and .
- Total indoor cooling heat load in and in Tons of Refrigeration () ().
- Required chilled-water mass flow rate () for a cooling coil temperature rise ().
1. Total Power, , and :
2. Indoor Heat Load in and : All electrical power dissipated inside the conditioned hall () becomes heat:
3. Required Chilled-Water Flow Rate ():
- , ,
- Indoor cooling load: ()
- Chilled-water flow rate:
6.2
A server rack cluster draws AC at power factor , powered by an online double-conversion UPS () with a battery string (, wiring/aging factor ). Determine:
- Apparent load power () and suitable UPS module sizing.
- Required DC battery power during a utility outage ().
- Minimum battery bank capacity ( at ) for a () backup bridge time.
1. Apparent Power and UPS Sizing:
Using two modular UPS units in an configuration () provides capacity with single-module failure tolerance.
2. DC Power Drawn from Battery:
3. Battery Bank Capacity for Bridge Time:
- Apparent load: (select two modules for )
- Battery DC discharge power:
- Minimum battery capacity at :
6.3
A server CPU executes a fixed workload of clock cycles. At nominal operating point , dynamic power is and constant static power is . Using DVFS, frequency and voltage are scaled to . Given , calculate the total execution energy before and after DVFS and the percentage energy saved.
1. Baseline Energy at :
2. Scaled Power and Energy at :
3. Percentage Energy Saved:
Baseline energy: ; DVFS energy: ; energy saved: .
Exercises
Practice Exercises — Energy Computation for ICT Infrastructure
A data center currently draws for IT equipment, for cooling, and for UPS/lighting losses.
- Compute its current and .
- After installing hot-aisle containment and economizers, drops to at the same IT load. How much annual electricity (, ) is saved?
1. Current and :
2. Annual Energy Saved at : New total facility power ; power saved .
- ,
- Annual energy saved:
Explain hot-aisle / cold-aisle containment, ASHRAE thermal guidelines, and economizer (free) cooling in data centers. Why does raising cold-aisle inlet temperatures from to reduce cooling electricity?
Airflow Containment & ASHRAE Guidelines: Hot-aisle/cold-aisle layout aligns server intakes toward cold supply aisles and exhausts into hot return aisles; physical containment doors and panels prevent hot exhaust from mixing with cold supply air. ASHRAE TC 9.9 recommends inlet air, safely supporting warmer supply setpoints.
Economizer Benefit of Higher Inlet Temperature: Raising supply temperature from to raises chiller Coefficient of Performance () and expands the number of hours per year when outside ambient air or water is cool enough to reject heat directly via air-side or water-side economizers, bypassing energy-intensive mechanical compressors.
Hot/cold-aisle containment stops hot and cold air mixing, ASHRAE sets safe inlet limits, and economizers use cool outdoor air/water directly. Higher inlet setpoints improve chiller COP and unlock far more compressor-free economizer hours.
Compare Offline, Line-Interactive, and Online Double-Conversion UPS topologies, and distinguish between and power redundancy architectures in ICT facilities.
UPS Topologies:
- Offline (Standby): Direct utility feed; switchover on outage; low cost for non-critical desktops.
- Line-Interactive: Adds automatic voltage regulation (AVR) for sags/surges with transfer; suits small network closets.
- Online Double-Conversion: Continuous isolation with transfer time and pure frequency/voltage conditioning; mandatory for data centers and core telecom nodes.
vs. Redundancy:
- : Adds one shared backup module to active modules; protects against a single module failure, but shares a single distribution path.
- : Duplicates the entire power distribution chain (two independent A+B feeds to dual-PSU servers), ensuring zero downtime even if an entire UPS, PDU, or utility feeder fails or undergoes maintenance.
Online double-conversion UPS delivers transfer and full power isolation compared to offline () and line-interactive () units. adds one backup module to a shared path, whereas provides two completely independent A+B power paths.
An enterprise server room runs underutilized physical servers, each drawing at CPU utilization. By virtualizing and consolidating workloads onto modern servers each drawing at utilization, calculate:
- The direct IT power reduction ().
- The total annual facility electricity saved (, ) assuming a constant .
1. IT Power Reduction:
- Legacy IT power:
- Consolidated IT power:
2. Total Annual Facility Energy Saved (): Each watt of IT load saved also eliminates of cooling and power-distribution overhead:
- Direct IT power saved:
- Total annual facility energy saved (including cooling/UPS overhead at ):
What is energy complexity in algorithm design? Describe four software and algorithmic strategies for energy-aware computing and green computing.
Energy Complexity: Unlike time complexity or space complexity , energy complexity quantifies the total joules consumed to execute an algorithm—capturing CPU arithmetic, cache/DRAM accesses, disk/SSD I/O, and network transmission costs.
Four Energy-Aware Software Strategies:
- Cache-Locality Optimization (Blocking): Minimizing DRAM and disk fetches, since off-chip DRAM reads consume more energy than ALU operations.
- Interrupt-Driven I/O & Sleep States: Replacing busy-wait polling loops with event-driven callbacks so CPUs enter deep ACPI C-states.
- DVFS & Race-to-Halt Scheduling: Scaling voltage/frequency down when tasks have deadline slack, or finishing rapidly to enter low-power sleep when leakage dominates.
- Precision & Compression Tuning: Using mixed-precision (
FP16/INT8) ML models and payload compression to cut computation and radio transmission energy.
Energy complexity measures total execution energy across compute, memory, and I/O. Key energy-aware strategies include cache-friendly data locality, interrupt-driven sleep instead of busy polling, DVFS/race-to-halt scheduling, and precision/compression optimization.