AWS has experienced exponential growth since its inception, growing from $0 to $500M in revenue in just 4 years. It releases new features doubling every year and has over 600,000 virtual machines running on just 80,000 physical servers across only 2 hardware vendors and 10 configurations. KT, a telecom provider, successfully converted to being a cloud service provider by building private, mobile, and public clouds based on best practices from real cloud competitors rather than traditional enterprise IT approaches. This allowed them to exceed expectations by outperforming AWS on pricing and provisioning times while reducing their capex spend by 79%.
2. Amazon Web Services:
• Zero to $500M/yr in 4 years;
• ~ 100% y/y growth
• Feature releases doubling every year;
• 400 employees (300 dev; 100 ops)
• 80,000 physical servers
• >600,000 virtual machines
• Only 2 hardware vendors
• <10 hardware configurations
• 10 price reductions in 18 months
(including free tier)
AWS features doubling every year. Amazon.com (e-tailer)
43 releases in 2009. • Sold >$1B in goods on mobile devices in
2008/2009
3. • >620M users
• +51%Y/Y growth
• >60,000 servers
• 550K+ apps
• 500MM+ downloads
• Over half of all mobile web pages
are Facebook pages
• Partnership with Skype means
that Telcos have a new voice and
messaging competitor
4. • New $500M data center
• 700,000+ square feet
• 112 containers
• 224,000 servers
• Runs with 45 people
8. Lesson: Scale requires new design patterns
Commodity hardware
Open source software
APIs for programmatic access
Simple systems
Horizontal scaling
Automation
Multi-tenancy
9. Lesson: To achieve scale build something
completely new
Personal Power
lowest efficiency Building Backup Utility Power Plant
4x more efficient 40x more efficient
SMB Computing
worst TCO Enterprise IT Google Data Center
4x more efficient 40x more efficient
10. Lesson: Cloud is not enterprise
Enterprise Infrastructure Cloud Infrastructure
Serves <500,000 employees Serves hundreds of millions of users
Expensive to build and scale 10x-40x more cost effective
Single-tenant Multi-tenanted
No API API to allow for programmatic access
Manual management Automated management
Manual provisioning Self-service provisioning
1 Admin: 50-100 servers 1 Admin: 1000-3000 servers
Hundreds of hardware configs 1-5 hardware configs for 100K servers
Designed to be “highly available” Assumes component failure, fault-tolerant
Enterprise “HA” h/w Inexpensive, reliable commodity h/w
Proprietary lock-in Maximum open source for less lock-in
Complex “vertical” stacks Simple “horizontally scaleable” services
Risk mitigation Risk acceptance
Must pre-build for peak usage Built to flex with usage
Great for steady state usage curves Great for elastic usage curves
Scale with bigger CPUs Scale by adding more CPUs
Design>Build>Dev>Operate DevOps = one discipline
12. Case study of a successful conversion
from service provider to CSP
13. KT decided to build 3 different clouds to
achieve all of its goals
Private Cloud Mobile Cloud Public Cloud
With these three clouds, KT intended to:
• Move 70% of internal IT to the cloud within its first operation
• Provide a low-cost platform for mobile-application deployment
• Provide public computing services at or below AWS pricing
14. KT chose to build their cloud like the real
cloud competitors rather than follow the
enterprise IT path
15. The result of this partnership exceeded all
expectations
CapEx Reduction Time to Provision
•3 clouds launched in less than
(5x) (50x) a year
•They outperform Amazon,
GoGrid, Rackspace*
•Public compute service is
priced 18% lower than AWS**
•Capex spend decreased by
79%
•Provisioning time was
improved 50x
18. In Cloud IT is a Weapon
Every choice directly affects service
profitability
•Return on invested capital
•Operating Expenses versus annuity
revenue streams
•Power utilization efficiency defines
max scale
•Automation reduces HR costs and
improves SLA adherence
19. CAPEX Based Market Pricing
$402/Month vBlock2 50% Margin
$201/Month vBlock2 Breakeven
$64/Month Amazon VM ECU
It is impossible to build a
$50/Month Arista/OS 50% Margin
profitable cloud with legacy
$25/Month Arista/OS Breakeven
enterprise technologies
20. Power Density = Profit Ceiling
Arista Network - 2.5% Everything you do after this
Cisco Network - 17% reduces the maximum
3MW Data Center
server capacity, thus
profitability
5000 Dual Socket Intel Westmere
4874 Dual@ 20 VMs per Server
Servers Socket Intel Westmere
Servers @ 20 VMs per Server
4125 Dual Socket Intel Westmere Choosing Arista, even with
Servers @ 20 VMs per Server
$5M/Month Revenue commodity compute and
$4.874M/Month Revenue storage increases revenue
$4.125M/Month Revenue potential by $8.98M/year
21. Arista Solutions - Cloud Starter Kit
‣ The 7124SX provides cost-effective 10GbE connectivity, at a density
perfect for the 2RU servers
Server 1 ‣ Because of the low density of this deployment servers are dual homed so
Cloud Mgmt no single-point-of-failure can impact the entire cluster - this model is
preserved until less than 20% of the capacity is lost on any switch outage
Server 2
‣ The 7124SX brings LANZ capability - so administrators can gain a comfort
level with the data flows, and see if congestion is impacting cluster
performance
‣ Scaling up to 20 Servers per cabinet maximizes network utilization and is
environmentally feasible in most 42U cabinets
Server 20
22. Arista Solutions - Scaling Out
‣ The 7050 is introduced as a spine switch - this design scales to 15 dual-homed cabinets.
‣ After 5 cabinets moving all cabinets except the Cloud Mgmt cabinet to single-homed increases the density of this design to 580
Servers (560 single-homed, 20 dual-homed)
Server 1
Server 1 Server 1
Cloud Mgmt
Server 2 Server 2 Server 2
Server 20 Server 20 Server 20
23. Arista Solutions - Continued Scaling Density
Server 1
Server 1 Server 1
Mgmt Node
Server 2 Server 2 Server 2
Server 20 Server 48 Server 48
25. We network all varieties of cloud
Community Cloud
Private Cloud
Mobile Cloud
Public Cloud Hybrid Cloud
26. and optimize the network for cloud providers
Automated
Extensible/Customizable
Scaled Management
Multi-Tenancy
Power/Environmental
CapEx Optimized
Hyper-Scale
27. Arista Cloud Networking
Public Cloud Architecture - Key Features
Arista EOS delivers key features necessary for
building scalable and profitable public cloud offerings
CloudVision - Management at Scale with common
open-standard /open-source message bus
Zero-Touch Provisioning - Fully automated deployment
based on HTTP that can auto-deploy switches and
servers in minutes
LANZ Latency and Congestion Monitoring - for high
performance cloud offerings actively monitor and
publish latency as a key performance indicator
Multi-Chassis Link Aggregation (MLAG) - active-active
uplinks and loop-free topologies. Use every link you
pay for to maximize performance and uptime
MLAG-ISSU - hitless upgrades from one version of
EOS to the next in MLAG topologies. Keep software
current and secure without impacting service levels.
28. Arista Networks: Cloud Portfolio
7500
7100SX Lossless, High
7048 7050
Density, Modular
Ultra Low Switching System
48-port Data 52/64 Port 10Gb Latency 24 Port supporting up to
Center Class Ethernet for Cloud 1/10Gb SFP+ 384 Wirespeed
Gigabit Ethernet Leaf and Spine Leaf Switches 10Gb Ports
Switch