Hardware Today: Meeting the Demand for On Demand

by Drew Robb

On-demand CPU schemes are all the rage. What do customers think of some new-to-market price schemes?

Why buy a CPU when you can rent one? Plenty of enterprise customers, it seems, are finding value in that logic. And many vendors, including Sun, HP, IBM, and Unisys, have brought products to market with a variety of pay-as-you-go arrangements for temporary CPU usage.

"We have a contractual agreement to receive and deliver data for customers, so we have to have capacity available to meet the peaks," says Mike Chester, senior manager of database administration at American Healthways, a disease management firm based in Nashville, Tenn. "Rather than having over capacity, we use on-demand CPUs to meet delivery demand and then turn it back down when traffic returns to normal."

To meet this need, American Healthways purchased two servers with HP Instant Capacity (iCAP) capabilities — an Integrity Superdome and an 8-processor Integrity RX7620.

The Superdome has three partitions: one for the Claim Utilization and Pharmacy system (CUPS), one for a data warehouse, and the other for testing.

CUPS uses 12 1.5-GHz Itanium 2 processors and has four additional iCAP processors on standby. The data warehouse has eight active processors and four iCAPs on hand. The testing group has four processors, which CUPS also uses on an as-needed basis. During weekly backups, for example, the company moves processors from testing to CUPS and transfers them back when the backup is complete.

"It is more economically feasible to buy this capacity than build-in over-capacity. It works out at about 80 percent cheaper per processor." — Mike Chester, senior manager of database administration, American Healthways

The capability to move processors around is particularly attractive to Chester. This means he can mix and match processors without touching the iCAP chips. When demand soars, however, he can turn to standby CPUs.

"It is more economically feasible to buy this capacity than build-in over-capacity," he says. "It works out at about 80 percent cheaper per processor.

American Healthways paid HP a license fee of 20 percent of the normal price for its iCAP processors. HP sells time in 30-day increments with the meter running only when the CPU is in use. The company initially bought 240 days. According to Chester, 30 days works out to be less than 5 percent of the total cost.

"This includes all the licenses for HP system software and support," says Chester. "Every year that you use iCAP, you are saving money, as you don't have to pay support and maintenance costs to have more hardware on the floor."

To turn the chips on, customers call or e-mail HP to obtain an access code, log on to the HP instant capacity portal, register, and then receive a user name and password. It takes about 15 minutes to register and get a password. From there, it is another 10 seconds to bring a chip into play. The next version of this program will come with five free temp days for proof of concept and disaster recovery testing. That means an organization can use it before registering. Further, if a regular processor is lost, an iCAP replacement automatically activates at no charge.

Other vendors have similar "pay-per-view" pricing models. Sun Microsystems has deployed such a model but with a twist. With the Sun Grid Utility, enterprises buy processor time on machines that reside elsewhere — no equipment purchase is necessary. No need to even buy a server with additional standby capacity &&151 simply buy the compute power and use it at you leisure.

According to Aisling MacRunnels, senior director of utility computing at Sun, the pricing is $1 per CPU Hour.

"If you use 100 CPUs for one hour, or one CPU for 100 hours, the price would be the same: $100," says MacRunnels. "If you turn on the CPU, for any length of time, you will incur a charge for it, and payment is collected in advance."

>> Rent to Own


As not every enterprise will opt for a pay-as-you-go model, some vendors are choosing alternate paths. Unisys, for example, rents CPUs for a certain amount of time, after which the organization owns them outright. Unlike HP and others, it does not charge a premium to have the CPUs sitting idle on the floor (i.e., charging for an 8-way unit even though only four processors will be used initially). The customer pays only for the processors that are turned on. Additional processors can be turned on, as needed and on a rental basis, initially. Unisys has set up the program so that the fourth time temporary processors are used, they receive permanent status without further payment. Ceilings can also be set on the number of consecutive temporary days a processor can be used to keep costs down.

"The Real-Time Capacity offering from Unisys provides a cost-effective way to scale up high-end servers and to rapidly respond to changes in the marketplace," says Radomir Zamurovic, CTO of Ruesch International, a provider of international payment and global currency services. "With this server, we can bring just the right capacity online exactly when we need it, and pay for the excess capacity when and if we use it."

The company bought two Unisys ES7000 510 servers, each with four processors active and four on demand. Ruesch first evaluated the various server vendors for the best server to host its database and database-centric applications. It selected Unisys, and then found out about the on-demand program.

"Real-Time Capacity was a no brainer, as you pay nothing to have more capacity right there," says Zamurovic. "Instead of paying for a half-empty box and then later bring it down and adding new stuff, I have a fully loaded machine I can activate by entering the key."

As these servers are collocated, Zamurovic says he would have added a lot of cost at the collocation site to add new processors — bringing the service down, physically adding more processors, then reconfiguring and rebooting. Now, it takes only a short time to simply activate the dormant chips.

Chip Consolidation

Other companies are turning to on-demand by coupling it with server consolidation as a way of coping with expansion. At GHY International, a customs brokerage firm based in Winnipeg, Canada, an IT staff of three deployed 14 consolidated servers (running Linux, Windows, and Domino) onto two IBM iSeries servers. After the first year, GHY estimates it saved nearly $100,000 from the switch.

Due to expansion, the company planned to add another nine Intel servers. But management nixed the idea of adding three entry-level staff members to support the three existing IT personnel. So GHY opted for two IBM iSeries 270 servers to supplement its iSeries 820. These machines run all core business applications as well as all Windows servers (via an xSeries server-in-a-card that slides into the IBM box. One iSeries also runs nine virtual Linux servers under the cover.

If you rent a processor and then use it endlessly, at some point it is less expensive to purchase it.
"We had additional CPU resources we can allocate on-demand to each of the various partitions," says Nigel Fortlage, vice president of IT at GHY International.

This came in particularly useful when GHY experienced a spike of 10,000 new e-mail messages in an hour from a worm attack. GHY had the CPU resources available to move to the logical partition on fly, and it was able to process good e-mail without performance loss.

According to Fortlage, he can now bring up and partition a new Linux server in 30 seconds. Thus, he has a fluid environment whereby he may go as low as seven virtual Linux servers and as high as 12.

"We can react and change our environment depending on what we want to do it," he says. "And we don't need to chase after budget to do it."

His iSeries has four CPUs available for standby usage. As of presstime, all four are active. They can be partitioned into one-tenth of a processor increments. Thus, although all four processors are being used, he still has a logical 1.3 processors available on demand.

But GHY won't be paying more for those 1.3 logical processing units. Fortlage explains that the decision to adopt an on-demand model was a functional rather than a financial one.

"We wanted the flexibility of having it all available, so we paid for it upfront," he says. "Putting a couple of processors on standby didn't work due to the flexibility of our environment."

Watch That Math

There is no doubt that on-demand computing has clear cut uses in today's computing landscape. Enterprises that want peak capacity or supercomputing resources on a temporary basis can benefit immensely. Similarly, those that like to change their partitioning infrastructure regularly can experience significant gain from an on-demand architecture.

But there are perhaps some cost pitfalls. If you rent a processor and then use it endlessly, at some point it is less expensive to purchase it. Fortlage, notes that after 40 days of rental, the amount spent is the equivalent of a processor purchase. He also cautions about 24-hour billing periods.

"You don't want to end up paying for 24 hours when you only used three," he says.

This article was originally published on Monday Aug 1st 2005
Mobile Site | Full Site