Cloud bursting for fun and profit…own the base, rent the spike.

eBay 2B page views/day 6000 application servers 9PB of data.

Cloud bursting: control costs and increase efficiency.

2000 servers provisioned at a fixed cost but only 300-1600 used at any given point of the day. A lot of “whitespace” or excess capacity.

How can we remove the peaks? Time shifting the workload helps a little bit, bringing us down to about 800 servers required.

Financial model structure: take the hourly demand in average TPS and max TPS per CU to get the hourly computing units required (CUs). Take the number of CUs in the internal cloud and the hourly cost, plus the hourly CUs in the extermal cloud, you get total hourly cost.

Cost benefit analysis shows that an external cloud at 4x the cost of the internal cloud still yields economic benefits in dealing with the spikes.

If the cost of the cloud is half of the internal cloud then it makes no sense to run anything internally.

In between is the interesting area where we need to make hard choices.

Cloud bursting is very interesting to eBay, allowing us to cut costs by at least 25%.

Lower the investment in infrastructure to focus on the business intelligence.

Cloud architecture is made up of many complex pieces. Compute, storage and network hardware, with a Cloud OS (IaaS control plane), and finally PaaS and SaaS layered on top.

Scale, automation, resource management, permission and policy management are the key challenges.

The hypervisor is NOT the Cloud OS, rather an essential component.

Large enterprises have shown that commodity hardware can lower costs. Design the application for commodity hardware and you can dramatically lower costs.

Network: topology no longer defines security, which now needs to be automatically and dynamically managed.

Federation: challenges are API, identity, data and application environment. Identity is the key.

Billing: metering, rating, and bill generation. The business is where the rating is.

Storage: Lots of data on enterprise platforms, opportunity to re-architect with low-cost storage. No one-size-fits all. The key is balance.

The cloud ecosystem has many components, many issues PER component. Focus on the key issues per component.

The cloud is coming and all that matters is cost. Rackspace cloud is for smaller clients, medium clients use hybrid strategies, and they are hoping large companies will use OpenStack.

Market will divide into 2 sectors: the self-service cloud, and the managed cloud. Just like in hosting versus managed hosting.

Service is what matters.

Pick the right platform for the right application. Rackspace’s strategy is that a hybrid approach of dedicated infrastructure and cloud is best served by a single provider.

SaaS – the “other” cloud. Rackspace hosts a huge number of SaaS platforms. SaaS is a huge agent of change in IT, and reduces IT staff.

OpenStack is a Nasa collaboration to provide an open-source platform that is hypervisor agnostic and provides storage portability.

Cedexis took a different approach to monitoring performance: inserted objects in all CDNs and convinced top sites to insert objects into their pages to gather real user performance experience.

15 Billion measurements in January, analyzed by Bitcurrent.

EC2 East outperforms Rackspace, Joyent, Google AppEngine and Azure.

If you deploy on multiple availability zones…do you use static geo-load balancing or base your decision on real-time performance information?

Turns out you get much better international coverage, a 20-30% improvement by using performance based load balancing.

Oriol introduces StarCraft, one of the most popular real time strategy games of all time, as a virtual world where you gather resources,

Starcraft is adversarial, long horizon, partially observable, realtime, concurrent. Very hard AI problem.

Challenge: Long Horizon, length of gae 10K to 100K frames versus chess with only hundreds of moves.

Designed Berkeley Overmind as a hierarchical search.

Realtime decisions, have to output actions every frame, 24+ times per second.

What we are learning with our AI efforts may be applicable to cloud computing problems…

Scott confesses that he’s a network engineer who loves hardware and datacenters…. He titled his talk “How I learned to stop worrying and love the cloud” Specific technologies work for specific solutions until needs change for a more efficient economical powerful technology comes along.

2006 Digg..so busy with the hardware side of the solution that they didn’t have time for application layer monitoring. Provisioning hardware took a disproportionate amount of time and effort. When they released more features they threw more servers at the problem. Lather, rinse repeat. Were not able to focus on the application which was really the important part of what they were doing.

Eventbrite moved from hosted servers to EC2 and Puppet. Single source of truth – master AMI, moved from Debian to Ubuntu. DNS/DHCP were major issues.Cloud computing is timesharing updated. Tips… use RAID 0 with EBS instead of RAID 10. Use 3rd parties for e-mail. Practice defense, redundancy and backups in depth.

Dreamworks is a CG film making company and a huge consumer of CPU. Cloud computing is a perfect fit. Cost effective scalable compute at their fingerprints.

A film takes 4-5 years, 200+ workstations, 50+ million compute hours, 100TB data, 500 million files, 10Gbps WAN, 40GBps LAN. Resource demands are bursty. On demands is a tremendous benefit.

In 2003 they used HP’s IaaS rendering farms for Shrek 2.

Getting the data to where the compute happens is a big challenge.

In 2010 3 CG feature releases in 3D, over 7M compute hours sent to IaaS. For 2011 they are increasing capacity 10x, adding a new IaaS provider, increasing network bandwidth 3x.

Redhat, KVM, MRG messaging queue management. Weblogic and JBoss middleware, and deltaCloud for cloud management.

Cloud computing allowed them to defer data center buildout, reduce cost, increase flexibility.

Want more multitenancy, more flexibility in ramping up and shedding infrastructure and better cloud storage.

Yahoo runs a “private cloud” of 400,000 servers. “Real-time” elasticity is the goal. What is elasticity – ability to dynamically provision computing resources to meet a business need.

The pants metaphor: Traditional approach to deal with expanding services…change the pants or loosen the belt buckle. But with maternity pants they expand to accomodate the growth on demand.

Yahoo! Cloud supports 600M+ users, 200PB of data, 100Bn events per day.

Spin up time for traffic spikes is a major problem, load shedding is the only current option. Turn off things that aren’t critical to accomodate the load.

What kind of pants are you wearing?