Post on 12-May-2015
© 2010 Cisco and/or its affiliates. All rights reserved. 1 © 2010 Cisco and/or its affiliates. All rights reserved. Cisco Confidential 1
Gary Berger Technical Leader, Engineering Office of the CTO
May 17, 2012
© 2010 Cisco and/or its affiliates. All rights reserved. 2
Technical Leader, Office of the CTO Data Center Business Unit
• 22 Years Infrastructure Architecture and Platform Development
• Performance and Capacity Planning • Data Center Design • Protocol Architecture • Application Design and Scalability • Software Defined Networking
@gbatcisco garyberger.net
© 2010 Cisco and/or its affiliates. All rights reserved. 3
• Partnering since 2008
• Advanced integration with Cisco Unified Compute System
• OpenStack Integration (Nova, Quantum)
• “Cloud in a Box” - High performance scaling to 1TB and 40 Cores.
© 2010 Cisco and/or its affiliates. All rights reserved. 4
1. Compute Intensive • Low number of tasks and small input size • This includes MPI workloads familiar in HPC applications.
2. Data Analytics • Larger data sizes familiar to Map/Reduce programming
model
3. Loosely Coupled • Modest data size but increasing the number of tasks • Indicative of data-grid applications and HTC which are
bounded by memory capacity but also can be bounded by local disk I/O
4. Data Intensive • Many tasks and large datasets. • Formidable challenge for networks with dense matrix • Categorized as Many Task Computing (MTC)
Data Size compared to Task Rate
AnalyticsData Intensive
Loosely Coupled
ComputeIntensive
1 1K 1M
Number of Tasks
Low
Med
High
Data Size
© 2010 Cisco and/or its affiliates. All rights reserved. 5
• Current Internet Trends
• Quick historical perspective and state of the “cloud”
• Data Center as a Business Archetypes
• Mechanical Sympathy
• Real World Challenges
• Service Centric Networking
© 2010 Cisco and/or its affiliates. All rights reserved. 6
• +150M Active Users • +340M Tweets per day
• 4B videos view/day • 800M visitors/mnth • 60H uploaded/min
• +900M Users • 3.2B Likes/Comments/day • +300M photos uploaded/day • 125B Friendships
© 2010 Cisco and/or its affiliates. All rights reserved. 7
0
2
4
6
8
10
12
2011 2012 2013 2014 2015 2016
Mobile Data Traffic (Exabytes/Month)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90%
100%
Operator A Operator B Operator C Operator D
Mobile Data Transfer Distribution
Other Web Video
Source: Cisco VNI Mobile 2012 Source: ByteMobile Mobile Analytics Report 2012
© 2010 Cisco and/or its affiliates. All rights reserved. 8
Unique problems that Cloudfy solves
© 2010 Cisco and/or its affiliates. All rights reserved. 9
Alan Turing
June 1912 - June 1954
© 2010 Cisco and/or its affiliates. All rights reserved. 10
• Time shared system
• Explicit control • Restricted scope • Tightly Coupled • Vertically
Integrated
Database Centric Client Centric Service Centric Host Centric
• Desktop applications
• Centralized File & Print
• Many dependencies • Low network
utilization
• Evolution of Client/Server
• 4GL Programming • Stored Procedures • Vertically Integrated • Proprietary
“Technical Debt”
• Loosely coupled components
• Web based interactions
• Almost Infinite Scalability
• Global scope • App driven
operational integrity
Web Centric
• Normalized Presentation Layer
• Ubiquitous Access • Ubiquitous API • Self-Described Data
“New Economy”
Sparse to Dense
© 2010 Cisco and/or its affiliates. All rights reserved. 11
© 2010 Cisco and/or its affiliates. All rights reserved. 12
ZCloud
© 2010 Cisco and/or its affiliates. All rights reserved. 13
Geographic Reach
New Sources Of Data
Capex Controls
Market Expansion
Your Business
Service Monetization
© 2010 Cisco and/or its affiliates. All rights reserved. 14
© 2010 Cisco and/or its affiliates. All rights reserved. 15
“Until now, cloud computing has been mostly about the distribution of applications” “The next wave of cloud computing will enable the sharing of the environment to run those applications.” “You will be able to take advantage of what we had to build in order to create those applications” Ben Fried, CIO Google 2012
© 2010 Cisco and/or its affiliates. All rights reserved. 16
© 2010 Cisco and/or its affiliates. All rights reserved. 17
Heterogeneous Multi-Tenant • Highly virtualized • Leverage compute arbitrage and
SPOT market • Benefits from a mixture of customer
market segments to randomize demand
• Complex engineering due to overlapping naming/addressing
• Complex operations due to uncoordinated modifications, interference due to competing access to shared resources
• Enhanced focus on security and isolation
Examples: Amazon EC2, Rackspace, etc..).
Homogenous Web Scale • Highly distributed • Leverages scale-out/parallel
application design • Minimizes heterogeneous applications
by providing higher level services and common resources management
• Enhanced focus on cost and efficiency due to large population.
• Operational separation of code, data, configuration and policy
Examples: Google, MSFT, Facebook, Yahoo
Unified Multi-Service • Highly flexible • Incorporates qualities of both HMT and
HWS • Purpose built to remove infrastructure
barriers to application development • Manages resources more efficiently by
controlling allocation via higher-level platform services
• Provides best ROI and flexibility through common abstraction libraries and runtimes
• “Its all about the app” • Operations as a Service
Examples: Amazon (DDB, EMR), RHEL OpenShift, MSFT Azure, VMForce
© 2010 Cisco and/or its affiliates. All rights reserved. 18
I/O Wall App Memory Wall
Having an understanding of the underlying architecture and behavior in order to build better systems.
Power Wall
© 2010 Cisco and/or its affiliates. All rights reserved. 19
Serialized Contention starts to dominate (i.e. locking)
Linear Growth (Scale-Up/In)
Coherency starts to force retrograde behavior O(N^2)
Amdahl
C(p) = p1+α (p −1)+ β p(p −1)
© 2010 Cisco and/or its affiliates. All rights reserved. 20
Presentation Tier
App Logic Data
Increased Delay/Limited Scalability
Network
Firewall
Load Balancer
Web
Firewall
Load Balancer
App
Firewall
Load Balancer
DBA Network
Network
Network
© 2010 Cisco and/or its affiliates. All rights reserved. 21
App Services
Caching &
Presentation
Data Services
Recipe
Cluster Manager
SDN Controller
© 2010 Cisco and/or its affiliates. All rights reserved. 22
application { name : myApp tenantID: tenantID service { compute { template: ucs_small_linux } network { template: publish_subscribe } storage { template: cache_persistant }
}
}
network{ name: publish_subscribe
qos: best_effort isolation: per_domain encryption: true msgPattern: pubsub
} storage {
name= cache_persistent cache { capacity: 5G
evictionPolicy: LRU } persistence{ block: 10TB file: extfs RAID: 10 }
}
© 2010 Cisco and/or its affiliates. All rights reserved. 23
• Effective Resource Sharing • Further away from the metal, the harder it is to understand (non-deterministic performance) • Contention grows while accessing shared resources • What instruments to collect analyze and model
• Programming Languages • Generally languages are insufficient for building large applications (lack of procedures in JAVA, lack of encapsulation in
Python, etc.) • Concurrency is still extremely difficult and hard to reason about (trend towards functional reactive programing) • Throw away code
• Network Scalability • Segmentation and Isolation • Address Learning • Application aware • Programmatic Interfaces
• Security • In-flight/At-Rest encryption • Proper tradeoff between performance and privacy • Rat-Hole because of lack of tools, developer education and highly incentivized and motivated hacker community
© 2010 Cisco and/or its affiliates. All rights reserved. 24
Thank you.