uc berkeley 1 beyond the hype: a berkeley view of cloud computing anthony d. joseph*, uc berkeley...
TRANSCRIPT
UC Berkeley
1
Beyond the Hype:A Berkeley View of Cloud Computing
Anthony D. Joseph*, UC BerkeleyReliable Adaptive Distributed Systems Lab
TNC 201031 May 2010
Image: John Curley http://www.flickr.com/photos/jay_que/1834540/
*Director, Intel Labs Berkeleyhttp://abovetheclouds.cs.berkeley.edu/
2
Outline
• What is it? • Why now?
– Cloud economics and opportunities• Challenges
3
Datacenter is new “server”
• “Program” == Web search, email, map/GIS, …• “Computer” == 1000’s computers, storage, network• Warehouse-sized facilities and workloads• New datacenter ideas (2007-2008): truck container (Sun),
floating (Google), datacenter-in-a-tent (Microsoft)• How to enable innovation in new services without first
building & capitalizing a large company?
photos: Sun Microsystems & datacenterknowledge.com
RAD Lab 5-year Mission
Enable 1 person to develop, deploy, operate next -generation Internet application
• Key enabling technology: Statistical machine learning– debugging, power management, performance prediction, ...
• Highly interdisciplinary faculty & students– PI’s: Patterson/Fox/Katz (systems/networks), Jordan (machine
learning), Joseph (systems/networks/security), Stoica (networks/P2P), Franklin (databases)
– 2 postdocs, ~30 PhD students, ~5 undergrads
4
Examples
• Predict performance of complex software system when demand is scaled up
• Automatically add/drop servers to fit demand, without violating SLO
• Distill millions of lines of log messages into an operator-friendly “decision tree” that pinpoints “unusual” incidents/conditions
• Recurring themes: – Cutting-edge SML methods work where simpler methods
have failed– Demonstrate applicability on at least 1000’s of machines!
5
Utility Computing Arrives• Amazon Elastic Compute Cloud (EC2)• “Compute unit” rental: $0.10-0.80 0.085-0.68/hour
– 1 CU ≈ 1.0-1.2 GHz 2007 AMD Opteron/Intel Xeon core
• No up-front cost, no contract, no minimum• Billing rounded to nearest hour (also regional, spot pricing)• New paradigm(!) for deploying services?, HPC?
Platform Units Memory Disk
Small - $0.10 $.085/hour 32-bit 1 1.7GB 160GB
Large - $0.40 $0.35/hour 64-bit 4 7.5GB 850GB – 2 spindles
X Large - $0.80 $0.68/hour 64-bit 8 15GB 1690GB – 4 spindles
High CPU Med - $0.20 $0.17 64-bit 5 1.7GB 350GB
High CPU Large - $0.80 $0.68 64-bit 20 7GB 1690GB
High Mem X Large - $0.50 64-bit 6.5 17.1GB 1690GB
High Mem XXL - $1.20 64-bit 13 34.2GB 1690GB
High Mem XXXL - $2.40 64-bit 26 68.4GB 1690GBNorthern VA cluster
• Major enabler for SW as a Service (SaaS) startups– Animoto traffic doubled every 12 hours for 3 days when
released as Facebook plug-in in April 2008
• Peak EC2 instances: – Mon 50, Tues 400, Wed 900, Friday 3400
Cloud as Major Enabler
But...
What is cloud computing, exactly?
8
“It’s nothing new”
“...we’ve redefined Cloud Computing to include everything that we already do... I don’t understand what we would do differently ... other than change the wording of some of our ads.”
Larry Ellison, CEO, Oracle (Wall Street Journal, Sept. 26, 2008)
9
“It’s a trap”
“It’s worse than stupidity: it’s marketing hype. Somebody is saying this is inevitable—and whenever you hear that, it’s very likely to be a set of businesses campaigning to make it true.”
Richard Stallman, Founder, Free Software Foundation (The Guardian, Sept. 29, 2008)
10
Above the Clouds:A Berkeley View of Cloud Computing
abovetheclouds.cs.berkeley.edu• White paper by RAD Lab PI’s and students
– Clarify terminology around Cloud Computing– Quantify comparison with conventional computing– Identify Cloud Computing challenges & opportunities
• Why can we offer new perspective?– Strong engagement with industry– Users of cloud computing in our own research and
teaching in last 3 years
• Goal: stimulate discussion on what’s really new– without resorting to weather analogies ad nauseam
11
Cloud Computing: True UtilityCloud Computing: App and Infrastructure over Internet
Software as a Service: Applications over the Internet
Utility Computing: “Pay-as-You-Go” Datacenter Hardware and Software
Three New Aspects to Cloud Computing
The Illusion of Infinite Computing Resources Available on Demand
The Elimination of an Upfront Commitment by Cloud Users
The Ability to Pay for Use of Computing Resources on a Short-Term Basis as Needed
Classifying Clouds
App Model for Utility ComputingSomething
New???
???
???
Amazon EC2
Close to Physical Hardware
User Controls Most of Stack
Hard to Auto Scale and Failover
Windows Azure
.NET and CLR… ASP.NET Support
More Constraints on User Stack
Auto Provisioning of Stateless App
Google AppEngineApp Specific Traditional
Web App Model
Constrained Stateless/Stateful Tiers
Auto Scaling and Auto High-Availability
Constraints on App Model Offer Tradeoffs… Lots of Ongoing Innovation…
Lower-level,Less managed“flexibility/portability”
Higher-level,More managed
“more built-in functionality”
• Instruction Set VM (Amazon EC2, 3Tera)• Managed runtime VM (Microsoft Azure)• Framework VM (Google AppEngine, Force.com)
14
Outline
• What is it? • Why now?
– Cloud economics and opportunities• Challenges
Why Now (not then)?
15
Economies of Scale for Humongous Datacenters(1,000’s to 10,000’s of commodity computers)
ElectricityPut Datacenters at Cheap Power
NetworkPut Datacenters on Main Trunks
OperationsStandardize and Automate Ops
HardwareContainerized
Low-Cost Servers
Technology Cost in Medium-Sized DC
Cost in Very Large DC Ratio
Network $95 per Mbit/sec/month $13 per Mbit/sec/Month 7.1
Storage $2.20 per GByte/month $0.40 per Gbyte/month 5.7
Administration ≈ 140 Servers / Administrator
> 1000 Servers / Administrator
7.1
James Hamilton, Internet Scale Service Efficiency, Large-Scale Distributed Systems and Middleware (LADIS) Workshop Sept‘08
Huge DCs 5-7X as Cost Effective as Medium-Scale DCs
16
Why Now (not then)?
• Common HW & SW platform– x86 as universal ISA, plus fast virtualization – Standard software stack, largely open source (LAMP)– Bet: Can statistically multiplex multiple instances onto
a single box without interference between instances
• Novel economic model: fine grain billing– Earlier examples: Sun, Intel Computing Services—
longer commitment, more $$$/hour
• Infrastructure software: eg Google FileSystem• Operational expertise: failover, DDoS, firewalls...• More pervasive broadband Internet
Unused resources
Cloud Economics 101
• Static provisioning for peak: wasteful, but necessary for SLO
“Statically provisioned” data center
“Virtual” data center in the cloud
Demand
Capacity
Time
Re
sou
rce
s
Demand
Capacity
TimeR
eso
urc
es
17
Risk of underutilization if peak predictions are too
optimistic – Wasted CapEx
Risks of underprovisioning
Lost revenue
Lost users
Re
sou
rce
s
Demand
Capacity
Time (days)1 2 3
Re
sou
rce
s
Demand
Capacity
Time (days)1 2 3
Re
sou
rce
s
Demand
Capacity
Time (days)1 2 3
18
New Scenarios Enabled by “Risk Transfer”
• Not (just) CapEx vs. OpEx!
• “Cost associativity”: 1,000 computers for 1 hour same price as 1 computer for 1,000 hours– Washington Post converted Hillary Clinton’s travel
documents to post on WWW <1 day after released– RAD Lab graduate students demonstrate improved
Hadoop (batch job) scheduler—on 1,000 servers
19
20
Outline
• What is it? • Why now?
– Cloud economics and opportunities• Challenges
Obstacles and OpportunitiesObstacle Opportunity
1 Availability of Service Use Multiple Cloud Providers; Use Elasticity to Prevent DDOS
2 Data Lock-In Standardized APIs; Compatible Software to Enable Surge Computing
3 Data Confidentiality and Auditability Deploy Encryption, VLANs, Firewalls; Geographical Data Storage
4 Data Transfer Bottlenecks FedExing Disks; Data Backup/Archival; Higher Bandwidth Switches
5 Performance Unpredictability Improved VM Support; Flash Memory; Gang Scheduling VMs
6 Scalable Storage Invent Scalable Store
7 Bugs in Large Distributed Systems Invent Debugger that Relies on Dist VMs
8 Scaling quickly Auto-Scaler; Snaphots for Conservation
9 Reputation Fate Sharing Reputation Guarding Services
10 Software Licensing Pay-for-Use Licenses; Bulk Use Sales
Open source reimplementations of Google AppEngine (AppScale), EC2 API (Eucalyptus), BigTable
(HyperTable)
Freedom OSS partnership with Amazon to allow FedEx-ing disks into their datacenters, Amazon hosting free public datasets to “attract” cycles
2/11/09: IBM WebSphere™ and other service-delivery software available on AWS with pay-as-you-go pricing
Cloud or Earthbound: “Should I Move to the Cloud?”
• Some app types made more compelling– Surge computing: overflow into the cloud– Extend desktop apps into cloud: Matlab,
Mathematica; soon productivity apps?– Batch processing to exploit cost associativity, e.g. for
business analytics
• Other apps more challenged– Bulk data movement expensive, slow– Jitter-sensitive apps (long-haul latency & transient
performance distortion due to virtualization)
• What about HPC/Grid apps?22
2008 Observed EC2 Topology
Caveat: only shows routers, not switches
EC2 Large Instance Network Performance
• Measurement:– Amount of data
transferred between 2 instances in 10 seconds
• Network quirks– 96% RTT < 1ms– Occasional route
changes lead to weird paths and increased latency
Effects on MPI Performance
• Many research opportunities to address issues– Collocated rack-level allocation schemes– Gang scheduling for clouds/virtual clouds– New network topologies/implementations
25[Walker, :login 2008-10]
Summary
• Many Cloud Computing Benefits:– Shift CapEx to OpEx , Scale OpEx to demand– Startups and prototyping, One-off tasks (Wash. Post)– Cost associativity– Research at scale
• Many Cloud Computing Challenges:– Availability – Data in cloud may be a “gravity well” ($$$ to move)
• Not ready for HPC applications yet– Opportunities to address performance issues
• More: http://abovetheclouds.cs.berkeley.edu/26
BACKUP SLIDES
28
Performance for Money Spent on EC2
• LINPACK cost effectiveness on EC2
29[Napper and Bientinesi]
Heterogeneity in Virtualized Environments
• VM technology isolates CPU and memory, but disk and network are shared– Full bandwidth when no contention– Equal shares when there is contention
• 2.5x performance difference
1 2 3 4 5 6 70
10
20
30
40
50
60
70
VMs on Physical Host
IO P
erf
orm
an
ce
pe
r V
M (
MB
/s)
EC2 small
instances
Power and Cooling Is Expensive!
Belady, C., “In the Data Center, Power and Cooling Costs More than IT Equipment it Supports”, Electronics Cooling Magazine (Feb 2007)
The Infrastructure for Power and Cooling Costs a LOT
Infrastructure PLUS Energy > Server Cost Since 2001
Infrastructure Alone > Server Cost Since 2004
Energy Alone > Server Cost Since 2008
Cost Effective to Discard Inefficient Servers
Power Savings Infrastructure Savings!
Like Airlines Retiring Fuel-Guzzling Airplanes
Willing to pay more $/server for more power efficient servers