For Those Playing the Home Version

You can follow along at

http://www.eecs.berkeley.edu/~jkuroda/talks/clouds/

Left/Right arrows, page up/down, clicking, and space bar

What I Will (try to) Tell You

• Some General Background
• What Is Cloud Computing?
• Where Did It Come From?
• What is good for?
• “Cloud Econ 101”
• Challenges and Opportunities
• Should I Cloudify?
• Real Experiences

General Background

• RAD Lab research Goals
• Why we might be on to something
  
  Research engagement with ...
  Google • Microsoft • Sun
  Amazon • Cisco • Cloudera • eBay • Facebook • Fujitsu
  HP • Intel • NetApp • SAP • VMWare • Yahoo! Research
• And in the interest of full disclosure …
  These companies do support us financially.

All Hype?

If you ask one notable industry personality, this is all just more marketing hype and buzzwords …

“...we've redefined Cloud Computing to include everything that we already do... I donʼt understand what we would do differently ... other than change the wording of some of our ads.”
--Oracle CEO Larry Ellison, Sept. 2008

There is a lot of hype, but we're here to cut through all that.

Definitions

• Software as a Service (SaaS)
• Utility Computing
• Public and Private Clouds
• Cloud Computing
• “The Cloud”
• Others as they come up

Defining SaaS

• Applications/Software Delivery over the Internet
• Name may be recent, but decades old idea
• Examples:

In some cases, applications are “just outsourced”, but some apps, like CRM for sales/support/marketing need to be mobile but also need to have a safe haven for various customer/client data - they couldn’t really exist as traditional apps.

Defining Utility Computing

• Again an old idea, but now it exists
• (Usually Believeable) illusion of infinite resources
• On-demand (order of minutes)
• Pay-As-You-Go, fine-grained billing, no upfront cost
• Examples:

Utility Computing provides the base on top of which various Software Services can be built and provided.

Defining Clouds

• Clouds
  • Instances of HW/SW Infrastructure
  • Provide the service called Utility Computing.
• Public Clouds:
  • available to external customers
  • have a pay-as-you-go billing model
• Private Clouds:
  • available only within an organization
  • billing may be obfuscated - “Overhead”

So What is Cloud Computing?

It is the combination of Public Clouds providing the service called Utility Computing on top of which various Software Services (Software as a Service) are built and deployed.

Someone (“Cloud Provider”) builds a Cloud (a whole other talk) to provide Utility Computing.

Customers (“Cloud Users”) build their own application on top of Cloud(s) to provide Software as a Service (SaaS) to end-users.

Defining “The Cloud”

Originated with the use of a drawing of a cloud to represent the telephone system and later network systems such as the Internet. Now a vague term used to abstract away any “complex system that takes care of something for me.”

I’ll use it here to mean “in the context of Cloud Computing”.

Where did all this come from?

What made all these old ideas possible in their present form?

• Big Internet Companies with Big Datacenters
  Huge Economies of Scale
• Widespread availability of broadband
  Everyone online, mobile devices
• Commoditization/Virtualization of x86 HW
• Maturation of common software stacks (free and $$)
• Ability to bill at fine-grained level

Not All Clouds Are Alike

• Type of Virtualization
  • Amazon EC2, 3Tera: Xen x86 VM, very DIY
  • Microsoft Azure: Bytecode (.NET) VM
  • Google AppEngine, Force.com, Cloudera: specific focus
  • Low-level/DIY versus High-level/more built-in management
• Storage
  • Amazon: S3 (http), Elastic Block Store, SimpleDB
  • MS Azure: MS SQL, Azure Storage
  • Google: BigTable (non-relational DB)

Cloud Econ - Provisioning

• Cloud User
  • Provisioning for peak loads - wasted capacity
  • Underprovisioning - lost business and eventually customers
  • Use Cloud, as load rises, request more Capacity
  • As load falls, return Capacity, pay only for what you use.
• Cloud Provider
  • Save Energy - turn off systems to match (load + buffer)
  • Turning systems off not as bad as once thought
  • Not just saving energy from systems but also from AC systems

Cloud Econ - When Is DIY Cheaper?

• A quick back of the envelope calculation assuming:
  • $.085/hr for EC2 Instance (32bit 1.7Ghz, 1.7GB, 160GB 2007)
  • $.076/hr to buy server ($2000/server, 3 yr depreciation)
• If peak load > than about 1.1 times avg, EC2 cheaper
  • This doesn’t factor in cost of Storage, Network, Power/AC, Staff.
  • Lots of ‘hidden costs’, hard to deconstruct

Benefits to Cloud Provider Utilize off-peak capacity, Sell SW (.NET), Defend a franchise, reuse existing infrastructure

Cloud Econ - Economies of Scale

Massive Economies of Scale - 5-7x factor [Hamilton 2008]

Cloud Econ - Location = Power

10,000 systems need a lot of power

Cloud Econ - Cap/Op Ex

“But what about Capital versus Operational Expenses!!??”

We get this one a lot

• In the long run, a dollar spent is a dollar spent.
  
  • A $ spent in highly utilized “Cloud Compute Time” or
  • A $ spent on moderately used system that’s obsolete in 3-5 yrs
• But yes, if your capacity needs are well understood,
  then DIY can be cheaper than Cloudifying

Cloud Econ - Risk Transfer

• Cost Assosciativity in Clouds
  • cost(1K servers x 1 hr) = cost (1 server x 1K hrs)
  • Washington Post: Hillary Clinton's travel docs processed <10 hrs after release of raw docs, posted to WWW within 26 hours.
  • RAD Lab - publish academic results using 1000+ servers
• Major enabler for SaaS startups
  • Animoto Facebook plugin - traffic 2x every 12 hours for 3 days
  • Scaled from 50 to 3500+ servers … then scaled back down

Cool! What else works there?

• Big/Huge one-off or periodic jobs
  • Batch Processing (example: Hadoop/MapReduce)
  • All that satellite/seismic/collider data
• Startups (or projects) with more need for service than HW
  • many “Web Apps” are deployed partially or wholly via EC2
• Prototyping
• Research at scale that can’t be done in-house

What else makes sense or are compelling?

• Desktop apps with large compute needs
  • CAD Tools, Matlab, SAS/S+
    Note: Matlab already can use EC2 as a compute backend
  • Photoshop, video/CGI processing
• Maybe more generic desktops apps? MS Office?
• Mobile applications
  • use mobile device as UI, compute/store “in the cloud”
  • mobile device becomes generic and easily replaced.
  • Early Example: Danger Hiptop (aka Sidekick), MobileMe

What doesn’t work so well … yet

• Anything requiring bulk data transfer. More on this later
• I/O bandwidth, low network latency guarantees
  • Memory/disk intensive work, no I/O bw guarantees available
  • HPC/Scientific computing with inter-thread communication
  • Amazon announced a Cluster Compute Instance - may help.
  • Fault/Delay-Tolerant, Low-communication methods under development
• Anything that needs real entropy

Challenges and Opportunities

Should I “Cloudify”

There are few parts to that question:

• Make use of some already available service “in the cloud”
  • GMail for Education
  • Salesforce
  • Online backup
• Move already existing application “into the cloud”
  (or develop new app in the cloud)
  • “Web Apps”
  • Database drive applications
• Build a (Private) Cloud and have it all in-house?

Building Clouds - Private versus Public

1. Doesn't factor in NRE costs to get app cloud-ified
2. Implies ability to meter & incentive to release idle resources

Should I build a Cloud?

• Why? Is cost savings expected?
  • Similar Economies of Scale unlikely for most
  • Beware double paying for bundled costs
• Internal incentive to release unused resources?
  • If not…don’t expect improved utilization
  • Implies ability to meter (technical) and charge (nontechnical)
• But consider Surge Computing

Surge Computing

• Easy: Use cloud for separate/one-off jobs?
• Harder: Provision steady state, overflow to cloud?
  • implies high degree of location independence, SW modularity
  • must overcome most Cloud obstacles
• Technical means exist
  • Eucalyptus - EC2-compatible (From UCSB) FOSS
  • AppScale - Google AppEngine-Compatible (also from UCSB
  • Microsoft Azure Appliance - run your own MS Azure instance
  • VMware vCloud?

In the Cloud, best practices are critical

• Authentication, data privacy/sensitivity
  • Data on public networks, stored in public infrastructure
  • Weakest link in security chain == ?
• Support/lifecycle costs vs. alternatives
  • Strong appliance market (e.g. spam filters)
  • Accountability gap for support

Bottom Line - all those things we should do in our own shops but don’t always do - we need to do them when we move our data into the cloud.

Real Life Experiences - Research

• Eucalyptus deployed on small (~40 node) cluster
• LOTS of Amazon EC2/S3 usage
• We can surge and migrate from one to the other
  • Same software, same VMs/AMIs
• Primarily for work that doesnt need Dept Infrastructure
  • Kerberos/LDAP, File Servers
  • Have to “package up” data/sw to use in the Cloud
  • License Servers
  • Work that relies upon these - open problem

Results - Research

• Higher Quality Research
  • Scale (100s to 1000s of systems) we can’t do in-house
• Faster Results -> we can solve new problems
  • Machine Learning/Data Mining research
  • eg - Near-realtime system log analysis
• Save money? … Wasn’t a primary gold
  • Goal was to increase scale of research
  • But, cost to get scale in-house - $1M
• Easy to quantify cost of research

Obstacles Encountered

• Accounting/Billing that rewards cost-effective cloud use
• Funding/Grant Agency culture hasn’t caught up yet
  • Funding for “Cluster” often can’t be used to buy time on EC2
• Tools require a lot of experience to use
  • But good opportunity for “appliances”
• Software Licensing Sucks
  • typically tied to user or fixed to CPUs/systems
  • we have a lot of experience dealing with this
    but typical user may not.

Education in the Cloud

• CS 98/198, Software as a Service
  • Started as Ruby on Rails class
  • gave them three cluster nodes to use - I was the sysadmin
• Moved it to Amazon AWS (EC2/S3)
  • More realistic environment, Can actually watch a DB fall over.
    Would have needed 200 servers for ~20 project teams
• VM image simplifies courseware distribution
  • Students can have root - OMG!
  • repair damage by re-instantiating image

Possibilities for Education

• No longer strictly limited by lack of in-house resources
• Class projects can take on life beyond end of semester
• Allows for wider collaboration - can be in other depts
• Students can provide more informed input on courseware

Summary - What’s new?

• New Scenarios are possible
  • Startups and prototypes without HW/capital investment
  • One-off/periodic tasks can exploit “cost-associativity”
  • Research/Education are otherwise unheard of scale
• Opportunity for cost savings/better utilization
  • Scale up and down to match usage
  • Economic motivation to scale down

Summary - Obstacles

• Dependency on 3rd parties
  • Data expensive to move, no universal format/method (yet)
  • Management interfaces not standardized (yet)
  • Still may rely on proprietery software
• Software licensing still sucks
• Security Considerations, IT Best Practices
• Hard to compare cost of Cloud vs in-house
• Administration/Accountability boudnaries

Summary - Cloudifying

• Save Money?
  • Economies of Scale unlikely, but can get better utilization
  • Need Resource Accounting/Incentive to Release Resources
  • Consider use of Cloud for “Surge” computing
• Even if you don’t move to Cloud, use as a driver
  • (Re)-examine adherence to best practices
  • Identify bundled/hidden costs - “overhead”

Conclusions

• Is this all a bunch of hype?
  • No.
• It this all a fad?
  • No.
• Is it for everyone?
  • Not yet, but that will change

Resources

• Reliable Adaptive Distributed Systems (RAD) Lab
• Blog for cloud discussion and our cloud computing whitepaper(Redirects to Blogspot)