Fabrics SDN and Big Data. What this means to your white
space
Carrie Higbie RCDD/NTS, CDCP, CDCS
Global Director Data Center Solutions and Services
TechTarget – Ask the Experts and columnist SearchNetworking, SearchVoIP, SearchDataCenters
BOD and Fmr. President BladeSystems Alliance Member Ethernet Alliance, IEEE
Columnist Performance Networking, SearchCIO, SearchMobile, ZeroDowntime, ComputerWorld,
IT World, Network World, Financial Review
Twitter: @carriehigbie
• TIA
• ISO/IEC
• Ethernet Alliance
• IEEE
• ANZ standards committees
• BICSI
• US Green Building Council
• Various World Green Initiatives
Technology Leadership
• Various World Green Initiatives
• Uptime Institute
• ASHRAE
• AFCOM
• 7x24Exchange
Cloud Security Issues
•
Data dispersal and international
privacy laws
–
EU Data Protection Directive and U.S.
Safe Harbor program
–
Exposure of data to foreign government
and data subpoenas
–
Data retention issues
•
Need for isolation management
•
Multi-tenancy
•
Logging challenges
•
Data ownership issues
Challenges for Cloud
• Infrastructures and data centers already built • Loss of control
• Shift of employment
• Potential loss of employment • Security – can be a big unknown • SLA’s
• Provider expertise in the cloud space • Portability
• Knowledge transfer • Compliance
• Location
• Lack of standards
• How to bill for services? • Public or private? Hybrid?
Driving Private Clouds
•
Portability
•
Confidence as a Service
•
What works in the cloud and
what doesn’t?
•
Vendor dependency
•
IT threat
•
Power users circumvent IT
•
HR involvement?
•
IT may not be involved in those
decisions
•
Tangible and intangible ROI
•
IT threat
•
Security
•
Bankruptcy
•
Geographic Diversity
calculations
•
Standards
•
Local regional sovereignty
requirements
Risk Assessment is Critical
• Type of information • Technology
– Vendor lock-in could be a threat – Open systems
• Vendors are trying to close their systems
• Location of information • Location of information
Lifecycle of a Data Center
• Inception • Design • Build • Operations • Retire/Retrofit • Retire/RetrofitInception
• We need a data center – Where?
– How big?
– How much power? – How much cooling? – Colo or not
– Cloud or not (public, private, hybrid) – Cloud or not (public, private, hybrid)
What can go wrong at inception?
• Dreaming too big • Ignoring new trends
• Living too cutting edge with nothing to fall back on • Relying too much on consultants previous designs • Relying on the wrong design build/consulting firms • Relying on the wrong design build/consulting firms
Example – What’s wrong here?
PDU’s add tocooling load
Walkway blocked
Open space mixes airflow
No access to cabinets – blocked
PUE
• Power Usage Effectiveness
• Defacto standard
• Total power / IT load
• Most data centers are in the 1.9-2.6 range
• The closer to 1 the better
• The closer to 1 the better
• Requires Intelligent PDU’s or some means to measure
USED power
• DCIE (Data Center Infrastructure Effectiveness) is
1/PUE to give percentage
Green Grid
•
The Green Grid is proposing the use of a new metric
•
Data center compute efficiency (DCcE)
•
and its underlying sub-metric, server compute efficiency (ScE).
•
These metrics will enable data center operators to determine the
efficiency of their compute resources, which allows them to identify
areas of inefficiency.
areas of inefficiency.
•
Using DCcE and ScE, data center operators can continually refine
and increase the efficiency of their compute resources in the same
way that they use power usage effectiveness (PUE) to improve data
center infrastructure.
•
CUE (Carbon Usage Effectiveness)
Ethernet, Energy Efficient Ethernet, Data Center
Ethernet
•
All things Ethernet!
•
EEE is a game changer, especially for 10GBASE-T
– Provides a true idle state – Significantly lowers power – Significantly lowers power
•
Data Center Ethernet collapses the backbone structure – layer 2
instead of layer 3
•
Idea is to increase speed (bridge not route) and provide priority
•
Ethernet has a LOT of overhead
Detailed Design
Original 2D customer floor plan concept
Customer Case Study
Thermal Analysis Overview
• Hot Aisle/Cold Aisle Configuration
– Based on customer-proposed type and placement of CRAC units, design create an optimal layout for thermal management, which was then run through thermal analysis software
Potential hotspots identified and corrected in the design
Theoretical analysis performed on multiple levels of equipment population to ensure ability to manage increased heat load of future equipment additions
Phase Phase 22 Phase Phase 22
ensure ability to manage increased heat load of future equipment additions
- Findings:
With airflow management practices such as:
• Vertical chimneys on SAN cabinets
Thermal analysis provided to user for communication with cooling vendor to ensure right-sized CRAC system
TIA 942, 942-2, 942-A
• This Standard (942-A) replaces ANSI/TIA-942
dated April 12, 2005 and its addenda.
• This Standard incorporates and refines the
technical content of:
– ANSI/TIA-942, Addendum 1, Data center coaxial
– ANSI/TIA-942, Addendum 1, Data center coaxial
cabling specifications and application
311 distances
312
– ANSI/TIA-942, Addendum 2, Additional guidelines for
Expanded Topology
Quick case studies
•
The savings behind a great
design
•
Must take into account the entire
data center ecosystem
•
Customer needs assessment
•
Knowledge of all DC products
•
Knowledge of all DC products
•
Knowledge of standards and
Traditional 3-Tier
•
3-Tier switch architectures
were common practice
– Core and SAN
switches in the MDA – Aggregation switches
in the MDA, IDA or HDA – Access switches in the HDA – Access switches in the HDA
•
Shorter distances
usually accommodated
more than two connectors
without exceeding loss
•
No longer adequate for non-blocking, low-latency communication
A better design=..
BASE-T Switches in HDA
Aggregate/Core
• Every red line is 24 or 48 channels of copper (primary network) • Every blue line is 24 or 48 channels of secondary copper
Top of Rack Switching 3 Tier
With ToR intercabinet connections change to
fibre
1, 3 and 5m limitation on passive,
10m on active SFP+
• In order to fill a 48 port switch,
you would need to be able to put
48 servers in the cabinet per
network
• Most data centers don’t allow
cabinet to other cabinet spaghetti
cabinet to other cabinet spaghetti
• Impossible in most data centers
– Limited by weight, power and
cooling
Installed port cost –
data center cabling media
10GBASE-CX4 INFINIBAND QDR FCoE FC/10GbE 10GBASE-Cu 5m 10GBASE-Cu Active 7m Hardware Cable Cost Maintenance 15% of
90 day warranty from electronics manufacturer
- 0 0 0 0 0 0 0 0 0 0 10GBASE-T (30m) 10GBASE-SR 10GBASE-LR 10GBASE-LRM 10GBASE-LX4 (switch to … 10GBASE-CX4 Maintenance 15% of cost YOY 20 YEAR WARRANTY
Costs for switching
List Total
Used
ports Unused
78Nexus 2000 32 port 10G $15,000.00 $1,170,000.00 1092 1404 78Redundant Power Supplies $500.00 $39,000.00
312SFP+ Uplink Ports $1,500.00 $468,000.00
Total for 2000's $1,677,000.00
10Nexus 5000 32 port $23,010.00 $230,100.00
39SFP+ Modules $5,200.00 $202,800.00
10Redundant Power Supplies $500.00 $5,000.00
Total for 5000's $437,900.00
Total for 5000's $437,900.00
2Cisco 7010 Core Switches $79,000.00 $158,000.00 2Redundant Power Supplies $7,500.00 $15,000.00 4Fiber Card for Uplinks $70,000.00 $280,000.00 $453,000.00
Grand Total $2,567,900.00 Does not include software
3 Years Maintenance $1,155,555.00
Zoned Approach
List Total
34 Nexus 2000 32 port 10G $15,000.00 $510,000.00 34 Redundant Power Supplies $500.00 $17,000.00 136 SFP+ Uplink Ports $1,500.00 $204,000.00
Total for 2000's $731,000.00 Savings $ 946,000.00 5 Nexus 5000 32 port $23,010.00 $115,050.00
20 SFP+ Modules $5,200.00 $104,000.00 5 Redundant Power Supplies $500.00 $2,500.00
Total for 5000's $221,550.00 Savings $ 216,260.00 2 Cisco 7010 Core Switches $79,000.00 $158,000.00
2 Cisco 7010 Core Switches $79,000.00 $158,000.00 2 Redundant Power Supplies $7,500.00 $15,000.00 2 Fiber Card for Uplinks $70,000.00 $140,000.00
$313,000.00 Savings $ 140,000.00 Grand Total $1,265,550.00 Savings $ 1,302,260.00 3 Years Maintenance $569,497.50 $ 586,057.50
Amount that could be used
for other equipment per
cabinet
• $48,420.71
• This is ONLY 39 cabinets!
• According to our ROI sheet 48 ports per cabinet
(24 each network) for 6A F/UTP and 30m
(24 each network) for 6A F/UTP and 30m
channels the cost is $105,399.00 or roughly 2 ½
cabinets worth of savings
• Total savings to customer
AFTER CABLING
Qty Description List Total
Network
Ports Used Unused
720 Nexus 2000 32 port $ 15,000.00 $10,800,000.00 23040 10080 12960 720 Redundant Power Supplies $ 500.00 $360,000.00
5760 SFP+ Uplink Ports $ 1,500.00 $8,640,000.00
Total for 2000's $19,800,000.00
Cost per port $859.38 $8,662,500.00 $11,137,500.00 192 Nexus 5000 32 port $ 23,010.00 $4,417,920.00 6144 2880 0 768 SFP+ Uplink Modules $ 5,200.00 $3,993,600.00
192 Redundant Power Supplies $ 500.00 $96,000.00
Total for 5000's $8,507,520.00
4 Cisco 7010 Core Switches $79,000.00 $316,000.00 4 Redundant Power Supplies $7,500.00 $30,000.00
24 Fiber Card for Uplinks $70,000.00 $1,680,000.00 768 768 0
Total for 7000's $2,026,000.00 Grand Total with top of rack $30,333,520.00
Port Count Total 29952 13728 12960
Nexus 2k's in top of rack
Required 5k's for uplinks
7k's and line cards
Port Count Total 29952 13728 12960
150 Nexus 2000 using required ports $ 15,000.00 $2,250,000.00 4800 4200 600 150 Redundant Power Supplies $ 500.00 $75,000.00
1200 SFP+ Uplink Ports $ 1,500.00 $1,800,000.00
Total for 2000's $4,125,000.00
Cost per port $859.38 $3,609,375.00 $515,625.00 19 Nexus 5000 32 port $ 23,010.00 $437,190.00 608 600 8 76 SFP+ Uplink Modules $ 5,200.00 $395,200.00
19 Redundant Power Supplies $ 500.00 $9,500.00
Total for 5000's $841,890.00
2 Cisco 7010 Core Switches $79,000.00 $158,000.00 2 Redundant Power Supplies $7,500.00 $15,000.00
3 Fiber Card for Uplinks $70,000.00 $210,000.00 96 76 8 1/2 the 7000's required
Zoned Approach - required ports 8400 at 14 servers dual network attached each
5k's required if centralized $25 million savings,
unused ports drop from 12,960 to 616
Enter New Switch Fabrics
•
Switch fabrics typically
only use one or two
tiers of switches
– Interconnection switches in the MDA or IDA
– Access switches in the HDA or EDA
HDA or EDA
– Aggregation switches are usually eliminated
•
Provides lower latency and greater bandwidth between any two
points
– Eliminates traffic having to travel through multiple switch layers
What this looks like- every server
connects to a more centralized switch
In the fabric world- also known as leaf and spine 2 Connector channel is basis for 40GBASE-T
Cross Connects in the Data Center
• Creates a convenient patching area
– Fiber panels that mirror switch ports connect via fixed links
– Creates an “any to all” configuration
– MACs made at the panels via fiber jumpers
• Can be located in a separate cabinet
– Keeps switches untouched and secure in their own cabinet
• Allows for one-time deployment • Allows for one-time deployment
of fiber from MDA to HDA
– Simplifies adding new equipment
• Fiber can be used for multiple purposes (networking or SAN)
– Swap a server’s fiber connection from networking to SAN using a jumper at the cross connect
Another View
• One-time deployment of fiber from cross connect in MDA to cross connects at End of Rows in HDA
• Fiber jumper
changes at End of Row
cross connects cross connects
allow fiber to be used for various purposes
• Each cross connect adds connection points and loss
Higher Speeds Have More Stringent
Insertion Loss Requirements
• Ethernet speeds migrating from 1 and 10 Gb/s to 40 and 100 Gb/s
• Fibre Channel speeds (SANs) migrating from 8 Gb/s to 16 and 32 Gb/s • Distance limitations and channel loss decrease as speeds have increase • Standards specify a maximum channel loss for each application
Fiber Type 1000BASE-SX (1 Gb/s) 10GBASE-SR (10 Gb/s) 40GBASE-SR4 (40 Gb/s) 100GBASE-SR10 (100 Gb/s)
Distance Channel Distance Channel Distance Channel
(m) Loss (dB) (m) Loss (dB) (m) Loss (dB)
OM3 1000 4.5 300 2.6 100 1.9
OM4 1100 4.8 400 2.9 150 1.5
Fiber Type
8 Gb/s Fibre Channel 16 Gb/s Fibre Channel 32 Gb/s Fibre Channel Distance (m) Channel Loss (dB) Distance (m) Channel Loss (dB) Distance (m) Channel Loss (dB) OM3 150 2.0 100 1.86 70 1.87
Or to put it another way=..
BASE-T Switches in HDA
Aggregate/Core
• If the top of rack switch is removed – the fiber in the top of the cabinet becomes a passive cross connect that can be used for SAN
• The same would hold true if the aggregate end of row switch was removed. This adds additional passive connections in the channel
Sacrificing copper increases fiber and $$==
• OM4 min recommended in TIA
Looking at costs
Three Tier Leaf/Spine DAC Leaf/Spine 10GBASE-T
Low density 14 Servers/Cab
Install Cost $ 8,816,885.18 $ 11,786,235.65 $ 8,638,321.02
Average Cost/Server Cab $ 61,228.37 $ 70,156.16 $ 59,988.34
Annual Power cost Networking $ 91,328.26 $ 101,419.78 $ 44,402.69
High Density 40 Servers/Cab High Density 40 Servers/Cab
Install Cost $ 16,295,362.88 $ 23,208,375.30 $ 18,722,843.39
Average Cost/Server Cab $ 113,162.24 $ 138,145.09 $ 130,019.75
Fabrics in particular
Leaf/Spine DAC Leaf/Spine 10GBASE-T
Low density 14 Servers/Cab
Total Equip/Cabling Cost $ 11,786,235.65 $ 8,638,321.02 Average Cost/Server Cab $ 70,156.16 $ 59,988.34 Annual Power cost Networking $ 101,419.78 $ 44,402.69 Total Cabling Cost $ 1,222,357.82 $ 70,327.30 Total Cabling Cost $ 1,222,357.82 $ 70,327.30 High Density 40 Servers/Cab
Total Equip/Cabling Cost $ 26,394,022.02 $ 21,596,114.19 Average Cost/Server Cab $ 157,107.27 $ 149,973.02 Annual Power cost Networking $ 177,610.75 $ 106,717.82 Total Cabling Cost $ 5,123,942.02 $ 2,078,260.76
Questions???
• @carriehigbie