Peaking Through the Clouds


structure_speaker_seriesCloud infrastructure services are particularly good at supporting variable demand and peaks with unpredictable timing or amplitude. Peaks are a challenge for CIOs, because forecasting too low may lead to poor performance or service unavailability, and guessing too high means paying for unneeded capacity. Peaking through clouds, instead of handling peaks with your own resources, can minimize cost while enhancing flexibility.

Whenever the peak-to-average ratio is greater than the premium (if any) associated with utility services, a pure cloud infrastructure approach will provide cost savings. In fact, even for more consistent demand, a hybrid data center/cloud solution may reduce total cost. However, given that a recent report showed that cloud services might cost twice as much as an enterprise do-it-yourself approach, a good rule of thumb would be that if your peak-to-average ratio is at least 2-to-1, then cloud infrastructure services should definitely be considered. For relatively flat demand, or when valleys can be filled with other workloads, cloud infrastructure may not provide cost savings, but other considerations may be relevant, e.g., response time improvement.

Since all spikes are not alike, I thought I’d use Amazon’s (s amzn) helpful Alexa web traffic monitoring service to look at some real-world examples. Alexa’s data collection methodology may not be foolproof, and using page views as a proxy for processing load may be imperfect, but both are good enough to characterize the kinds of demand curves commonly found.

Relatively Smooth Demand: For large sites with big customer bases that provide daily value, demand can be astonishingly stable and predictable. For example, people run searches every day, check their email every day, and network socially every day — whether at work or at home. Moreover, search providers can fill in demand gaps with extra web crawling to maximize utilization.

Periodic Peaks: (s wmt) also has predictable variation — despite the recession — with a doubling of traffic from Christmas shoppers and a quadrupling of traffic on Black Friday/Cyber Monday. Given this 4-to-1 ratio, should use cloud infrastructure services, which it does.


Irregular Peaks: Apple (s aapl) has a 2-to-1 peak-to-average. It’s pretty easy to pick out the original iPhone price cut, the iPhone 3G and 2.0 software release, and, just recently, the 3.0 downloads. While a CIO might want to smooth these spikes out to enhance utilization and reduce infrastructure costs, part of what makes Apple such an iconic company is the way it builds excitement around product launches. Its peak-to-average ratio is right at the breakeven point, but since many of these spikes are large software downloads, cloud services (for content delivery) make sense.


One-Time Events: China spent a reported $44 billion for the Beijing Olympics, a feat unlikely to be rivaled anytime soon. While the timing was never in question (08/08/08), and clearly there was broad interest across a global audience, how exactly could one predict demand variation for, say, any particular country’s team and their associated web site? Some team sites had a 25-fold change from August to September.

Unforeseeable Events: These have unpredictable timing and amplitude. A good example might be market meltdowns, 9/11 or Katrina. News sites have no way to predict the next hurricane, global nuclear escalation, or celebrity scandal.

And content distribution and syndication can create a “flash flood” at the origin site. For example, Yahoo! (s yhoo) sends 25 million referrals (i.e., clickthroughs) a month to members of its Yahoo! Newspaper Consortium. A single headline link at Yahoo! about puppies saving a freezing boy lost in the woods drove nearly a million referrals to the New York Daily News last December.

Deadlines, Openings and Windows: A deadline, such as for tax filing or DTV transition, can drive peaks — similarly, an opening or release, such as a product launch or when tickets go on sale for a concert or presidential inauguration. Limited availability can be a driver, but so can the need for instant gratification, via a new game or application download. When there are both a start time and a deadline, a window of opportunity (e.g., a weekend sale, an open enrollment period for benefits, or product testing and quality assurance) exists.

Tax filing illustrates all of these nicely. There is a window between early January and April 15, delineated by an opening where the prior year’s financial information becomes available and the filing deadline. There is a wide peak from January to March, as some taxpayers, who are either more organized or expect a refund, do their filing. Then, there is a sharp peak as those who owe money or are otherwise procrastinators, wait until the last minute. The number of such filers can be hard to predict.


While enterprises can do their best to smooth workload by aggregating demand from different business units and filling valleys with deferrable and batch jobs, spikes are a natural part of life’s ups and downs. What does your demand profile look like?

Joe Weinman is Strategy and Business Development VP for AT&T Business Solutions.




I think this is a nice take on using clouds (really I think your looking hard at CDNs here) for font end demand. Certainly the MJ incident yesterday is a great example of peaking and some sites without reliable fail over CDN partners slowed or went down.

Where the simplicity of the fail over model will really be, and is already being, tested is the real time web. Many of the site examples you give here the demand is mostly on read operations for content, not writes. Twitter would no doubt love to fail over to a CDN, but CDN’s don’t do well with write operations, and can’t due to data-coherency issues on the back end.

Net, net, the nice segmentation and fail over you speak of here is fine for most applications, but real time aps face a bigger challenges that may require a full fledged cloud development model.

Thanks for being really funny all day yesterday!

James @wattersjames

David Deans @ BTR

Joe, it’s interesting to compare the common “managed cloud service” capabilities between AT&T, Verizon and BT. Some large multinational companies may choose to use all three providers.

However, based on your observations, do you now see an opportunity to deliver distinctive new pre-sales consulting services to enterprise CIOs?

As an example, could a “predictive analytics” modeling tool aid the CIO decision making process of on-demand cloud service procurement. Also, we’ve used Erlang formulas in the networking arena to determine load balancing, do you see the evolution of similar formulas in the emerging cloud services space?

Moreover, is there demand for a dashboard-like application that would enable a CIO to use historical data from a hybrid solution to predict upcoming peaks — and then purchase “just in time” capacity from pre-approved Service Providers that offers the best deal?

The whole “supply vs. demand” — and augmentation balancing — topic seems to be a place where meaningful differentiation can emerge in the marketplace.

David Deans
Business Technology Roundtable

Comments are closed.