Data Centre Design and Operational

Best Practices

Ed Ansett
Managing Partner
i3 Solutions Group

ABB
5th July 2013
Causes of Failure in Data Centres

FAILURE WHO? WHAT?

System Design Consulting Engineer Human

Commissioning or Test Deficiency Commissioning Engineer Human

Equipment Design OEM Human

Equipment Failure No one Machine

Operator Error Ops Human

Installation Error Contractor Human

Maintenance Oversight Ops Human

Natural Disaster Act of God Various

High Profile Data Center Outage

Hot summer day, utility power outage, data center at full load 7.2MW
Four 2.5 MW generators installed (N+1 configured)
One generator fails to start (now N configured)
Running on 3 generators
30 minutes later another generator fails (now N-1 configured)
Now 5MW capacity supporting a 7.2MW load
Remaining generators overload 60 seconds
Cooling plant has no power
IT equipment begins shutting down (over-temperature) Why did they
have to lose the
DC data centre runs on UPS for another 30mins (2N 15mins each side) whole data
Total data center failure 30 minutes later centre?
Utility restored after 6 hours Why couldnt
Data Center fully restored after 8 hours later the data centre
load be reduced
Numerous senior managers fired from 7.2MW to
Worldwide enquiry launched 5MW? There
Litigation and financial penalties was enough
time.
Highly publicized therefore reputational damage
Data Centre Outage Root Cause
Data centre operations under pressure to keep costs to a minimum

FM tendered and outsourced to the lowest of three bidders

FM team were technically skilled but did not receive adequate transitional training from
previous contractor or DC owner

Incident occurs 3 months after new FM contractor appointment

DC owner decided to scale back maintenance from pro-active to reactive

Pneumatics systems not fully maintained or regularly tested

On the day the Gen No1 failed to start due to a blown seal in the Pneumatic System 1

Ops didnt know how to manually re-route the pneumatic system to start Gen 1
Data centre failure was inevitable

Huge negative impact on the DC owner

Data centre CIO tasks us with an independent no-holds barred review

What was the root cause, who was responsible and why?
The Answers are in the Universal Learning Curve

Depth of experience is a combination of relevant working experience,

education (cognitive knowledge), skills (psychomotor) and effective learning,
teamwork, awareness, knowledge sharing, communication etc.
 Why is 100% Uptime Impossible

All mechanical, electrical and electronic components are characterised a Mean Time
Between Failure, typically shaped like a bathtub
Failure Rate

The Bathtub Curve

Constant Failure Rate

Time
Operational Best Practices

Plan for Failure

Manage Capacity
Standard, Maintenance and Emergency Operating Procedures
Label Everything
Staff Levels
Training
Record Management
Communications
Human Error

There is no such thing as 100% uptime

Analyses of space, nuclear, aviation, chemical and other industries and reports that 80% of
all failures are down to human error
This correlates with the uptime institutes reports of approximately 70% of data centre
failures attributed to human error
Human failure rate is explained by the Universal Learning Curve which is an exponential
curve that varies in terms of organizational experience and operators depth of experience.
Gradually we all become complacent and therefore it is better to plan for an inevitable rare
failure (100,000 to 200,000 hours accumulated experience), than to assume / believe that
the failure will never happen.
Work in the Data Centre

All operations need peer review, rehearsals and full switching plans

Also need approval of the business or customers to carry out the work

Program of work to be carried out

Time frame

Risk review

Roll Back plans i.e. when to Roll Back of program to stay within operational
widows
Switching plans

Switching schedule of all steps

Tags fixed to each breaker/switch with breaker/switch reference number

and switching plan step number
Operating Procedures (SOPs and MOPs)

Standard SOPs (Standard Operational Procedures)

Procedure will generally have prior planning and approval required before operating
and implementing and may cover multiple devices and equipment

e.g. Software upgrades such as UPS firmware promotion

Planned MOPs (Maintenance Operating Procedures)

As the SOP but written specifically to cover a maintenance activity

e.g. Put UPS into bypass for invertor maintenance

Permit to work

Also included would be electrical Single Line Diagram and switching tags
Emergency Operating Procedures (EOPs)
Absolute minimum procedure required to safely and reliably operate the equipment in a
fast response incident
Example EOP to open and close an ACB (Air Circuit Breaker) rated at 4000A
Data Centre Infrastructure Management

Capacity planning, forecasting, Optimization, operational

simulation, analytics BI, load management

Data management, integration and reporting

IT Services and Systems

Management, VM Mgmt Power
Environmental management,
monitoring
power
and reporting capping

Asset,
configuration Building Life,
Cooling
and change Safety, Security Power, energy
control,
management measuring,
BMS,
modeling
alarms etc.

Data collection, meters, sensors

Data Centre Design Begins with Understanding

Business Processes: Functional elements

Application Chain: Applications , Application Management
IT Infrastructure: Compute, Network, Storage, VMs and Ancillaries
MEP Infrastructure: Power, Cooling and Ancillaries

Business Process Application Chain Technology MEP Infrastructure

1 x Infrastructure a e.g. PDU 5 & 6

Business Process Application Chain Technology MEP Infrastructure

2 y Infrastructure b e.g. PDU 7 & 8

Business Process Application Chain Technology MEP Infrastructure

3 z Infrastructure a+b e.g. PDU 5,6,7 & 8
Design Best Practice

How is application resilience implemented (onsite clusters, async/sync mirroring)

How is technology infrastructure resilience implemented (network/compute/storage)
What SLAs are required of the data centre
Unless we can answer these questions everything else in M&E design is guesswork
Then how skilled are the operations team in implementing EOPs
Now we can start the M&E and building design:
How big, what capacity, what density, growth plan, how many dc modules
Watts per square meter vs. kW per rack
Business SLAs determine the design
Design determines the budget
Business TCO balance and CSR drives PUE

All this before weve decided what to build and where to build it!!
 Cost Optimization Requires Reliability and Availability Analysis

99.9999 Adding redundant

components improves
availability and is expensive.
99.999
Availability

Adding too much redundancy

99.99 eventually causes availability
to reduce due to complexity
99.9
Where are you on the curve?
99.0

9.0

Cost $
Reliability Modelling
Using Probabilistic Risk Assessment
Why do Reliability Modelling?

Quantifies the risk of failure

Identifies vulnerabilities
Utility
Alignment of business mission and performance

Create action plan for site hardening Xfmr

Performance benchmarking
ACB
Support business case expenditure
MSB
What is involved?
MCB MCB
Step 1: Develop Resiliency Metrics and Quantify Reliability
Expectations

Step 2: Develop RBD Model and Evaluate Redundancy

Step 3: Perform Reliability Modelling
CL CL
Step 4: Upgrade/Alteration Recommendations and Performance
Benchmark Findings
Results

Reliability Data Sources

The reliability data sources used are from IEEE Standard 493-1997 Gold Book and Reliability
Analysis Center NPRD-95

Use of manufacturers data is avoided due to the impartial figures quoted

Now are we Ready to Decide

Design topology (concurrently maintainable, fault tolerant, hybrid etc.)

Module sizes (250kW, 500kW, 1000kW etc, 200sqm, 400sqm, 600sqm etc.)
Module types (fixed or mixed density, CRAU, In Rack, In Row etc.)
Type, size and configuration of cooling (Water or air cooled chiller, economizers, DX)
Type, size and configuration of power (block redundant, distributed redundant , 2N etc.)
Type of UPS (fuel cell, rotary, static)
Raised floor or no raised floor
AC or DC power
Reserves (power, water and gas)
Density (3, 5, 7, 10, 15, 20, 25, 30 kW per rack etc.)
Single or hybrid tier system
DCIM, BMS, SCADA control
Common Design Errors

Design with the operator in mind avoid multiple bypasses and other complex operations
The most unreliable component in the data centre is always batteries
Followed by anything that has mechanical moving parts (chillers, pumps, relays etc.)
Anything more than 5 minutes of battery autonomy is a waste
Hot & Cold aisles are wasteful without containment or precision airflow
Data centres comprises business service lines of varying priority, design accordingly
Mirrored tier data centres are orders of magnitude more reliable than a single data centers
When looking at PUE, consider partial load, its more important than full load
Tier 4 can be justified where the frequency of generator ops is high due to a weak utility
The tier system is a guide, nothing more
The key data centre design standard is now ANSI BICSI 002 2011
Cooling economizers do not provide good ROI in the tropics (Sensible PUE target 1.4)
There are new cooling technologies that will radically improve tropical PUE to around 1.15
Dynamic workload allocation will force data centres to be more responsive to applicatons.
Key Design Rules

Understand the IT strategy, engage with your IT group, understand their constraints

Dont add unnecessary redundant paths - complexity causes errors

MEP, IT infrastructure and application chains are aligned

The design is scalable, businesses need to be agile, so does the DC

The design is flexible, experience shows the future isnt always predictable

Design for operations, remember the easier it is to operate, the better it is

Dont overprovision its easy to do, its wrong, dont do it

The business SLA requirements are met, all DC stakeholders need to be aligned
IT StackAlignment

C-Level prioritize service lines (highest to lowest)

Identify application dependencies

Map applications to physical devices dedicated and virtualized

Configure VMs based upon C-Level category priorities

Avoid mixed priorities on VMs where possible

Locate VMs and dedicated platforms physically based on priority levels

Configure Cooling to service line priorities

Configure Power to service line priorities

Locate Common Priorities Together
Dedicated Servers and VMs
Increasing Priority (P1 highest)
P1 P1 P2 P2 P2 P2 P2 P2 P3 P3 P3 P3 P4 P4 P4 P4

A Side
B Side
P1 P1 P2 P2 P2 P2 P2 P2 P3 P3 P3 P3 P4 P4 P4 P4
Data Centre: Prioritized Cooling

P1 P1 P2 P2 P2 P2 P2 P2 P3 P3 P3 P3 P4 P4 P4 P4

A Side
B Side
P1 P1 P2 P2 P2 P2 P2 P2 P3 P3 P3 P3 P4 P4 P4 P4
Data Centre: Prioritized Power

P1 P1 P2 P2 P2 P2 P2 P2 P3 P3 P3 P3 P4 P4 P4 P4

A Side
B Side
P1 P1 P2 P2 P2 P2 P2 P2 P3 P3 P3 P3 P4 P4 P4 P4
Thank You

Ed Ansett
Managing Partner
i3 Solutions Group