Riding the peaks: Market-based systems help manage demand

by Simon Firth

Airlines, utilities, hotels, Internet service providers: These are vastly different industries with vastly different product offerings, yet all share a key problem: managing peak demand.

In each case, customers typically want more computer processing power, or tickets, or electricity at some times of the year (or day, or week, or month) than others.

Some of these times can be predicted, but some can't. So suppliers worry about having adequate resources to meet peak demand. They'd like to be able to plan ahead.

For customers, the problem is reverse – will the resources they need be available at peak periods and, if so, at how high a price?

Both would be a lot happier, says HP Senior Fellow Bernardo Huberman, if it was possible to structure prices so that demand for any particular product evens out.

That can be done, Huberman argues.

To accomplish it, he says, "one needs to design new mechanisms for reservations that induce users to reveal how likely it is that they will use the resources they reserve."

Costs of current systems

A reservations system allows people to pay up front to keep a resource (say an airline ticket, or computing power in a data center) available to them at some point in the future. The price they pay to hold that resource for peak times is less than if they bought it at the time they actually needed it.

“But the problem with reservation systems,” says Huberman, “is, if they are not structured well, people can strategically 'game' them.” People can lie about how likely they are to use a resource and get a better price than if they revealed their true intentions.

Or worse, they could not use it at all, with the consequent cost to the provider, who cannot sell that reservation to someone else.

Market-based resource allocation

Huberman, who directs the HP Information Dynamics Lab, and researchers Fang Wu and Li Zhang recently published a new model for a market-based resource allocation system that they believe is ‘game-proof’.

The inspiration for their work was the desire to manage peaks in demand for IT resources. But they believe their model could be applied to any industry that faces the same problem of uneven demand and thus requires a reservation system.

“What we came up with,” says Huberman, “is what we call 'truth-telling' reservations. It’s a pricing structure for reservations that forces people to reveal the true likelihood of their using that resource in the future.”

Honesty pays

Under such a structure, the more likely you are to use whatever you are reserving, the less you pay for the reservation as a non-refundable down payment. But there is also a sliding scale of penalties for not using the resource. The more likely you said it was that you would use the resource (for which you got a price break) the more you pay for canceling.

The two scales set it up so you will always pay less by being honest about your likelihood of using the resource than if you either wait and end up having to pay peak prices, or if you lie to get a price break and then have to pay a higher penalty when you don’t use it.

Under this kind of system, argues Huberman, buyers get a better price for covering their future -- but still uncertain -- resource requirements, while suppliers get predictable demand. An online merchant that reserves IT time, for example, gets assurance that it will have the computing capacity it requires to meet its own customers' needs during peak periods.

The result: peak-time reservation prices are lower for the buyer, and the supplier has an accurate (and therefore profitable) way of knowing what demand is going to be.

Predicting demand

For such a model to work well, however, buyers need to be able to realistically predict what their demand for any particular resource will be days, weeks, or even months ahead.

“We have a simulator tool that will do that,” says Huberman. “It allows you to measure past usage and predict how you would price in the future.”

This simulator, designed for data center customers, estimates the cost of new reservations from a set of historical customer data. It allows users to adjust the amount of CPU power, memory and bandwidth they might need, and then be either more aggressive or conservative about future demand.

All this, Huberman argues, provides “a powerful ‘what-if’ capability to both the resource provider and the customer for estimating outright costs and risks associated with fluctuations in customer demand.”

Possible applications

The HP team is hoping to trial their new reservation model soon within HP Labs.

One candidate for such testing is the Labs’ new virtualized market system for allocating resources in distributed computer clusters.

With this system, a customer pays for usage in a kind of resource spot market. This allocates the cluster more efficiently than the more common time-sharing model, and it allows users to change allocations in seconds. Adding ‘truth-telling’ reservations should only improve the resource-allocation system’s power to efficiently allocate IT resources.

Huberman also hopes the model will find broader use.

For many industries, he says, the issue of reservations is huge.
Airplane tickets, hotel room reservations, even the financial services, all need an efficient reservation system to manage peak demand, he notes.

The power of markets

If Huberman, Wu and Zhang’s ideas are widely adopted, expect changes to hit even the humble office cubicle.

“Today we are over-provisioning IT,” notes Huberman. “I have a desktop. It’s incredibly powerful, but half the time it’s not being used.

“It’s the same as the difference between having access to any airplane for free and having to pay a ticket when you really need to fly,” he explains, suggesting we should all have budgets for the compute time we use at work.

“Markets are good,” Huberman adds, “because they force you to express your preference by paying a price.”