Cloud Economics Favor the Small Workload
Cloud computing can cut the cost of running small workloads though may actually be more expensive for larger compute jobs, compared to the costs of running such work in-house, according to researchers from Pennsylvania State University.
Researcher Byung Chul presented these findings at the Usenix "HotCloud 2011" Workshop on Hot Topics in Cloud Computing, held this week in Portland, Oregon.
Chul, along with Bhuvan Urgaonkar and Anand Sivasubramaniam, analyzed the costs of cloud computing in the paper, "To Move or Not to Move: The Economics of Cloud Computing," which was submitted to the conference.
"Many people expect cost savings compared to traditional hosting for the reasons of the pay-as-you go nature of the cloud, or its elasticity. But there is no consensus about whether the cloud really provides the economic benefits or not," Chul said. "The goal of our study [was] to systemically investigate what conditions or variables affect the costs in what ways," Chul said.
Purveyors of cloud services like to tout that moving workloads to the cloud can save organizations a lot of money. At the very least it eliminates the capital expenditures needed for equipment and software. And cloud services also, in theory, can run equipment more efficiently, and hence less expensively, than organizations themselves can.
The benefits are not quite so obvious, the researchers argued. Cloud savings depend on many factors, such as the workload intensity, how much the application will be used in the years to come, how much storage is needed and the software licensing costs.
For the study, the researchers constructed two sample multitier workloads, based on the Transaction Processing Performance Council's TPC-W benchmark, which emulates the workload of an online book store, and TPC-E benchmark, which emulates online transaction processing in a brokerage firm.
The researchers then calculated out the costs of running these workloads over a 10-year time period, either in-house, in an Amazon EC2 (Elastic Cloud Compute) or Microsoft Azure hosted cloud service, or as a combination of in-house and these cloud services, called a "hybrid cloud."
"Overall, we find that in-house provisioning is cost-effective for medium to large workloads, whereas cloud-based options suit small workloads," the paper said.
Small workloads were defined as those with 20 transactions per second, with a growth rate of about 20 percent per year. Such workloads would cost about US$10,000 a year to execute in-house, while costing only $1,000 a year in a cloud. Over the period of 10 years, however, the costs of the two choices may even out, thanks to the 20 percent growth rate of the workload.
The tests assumed that hardware and software would be refreshed every four years. Thanks to Moore's Law, the new hardware would have improved performance, which could help keep costs down for in-house implementations. Cloud computing pricing, on the other hand, may not reflect the lower costs enjoyed by the newer hardware, the researchers observed.
For large workloads, the cost benefits of running in the cloud could run higher than keeping them in-house. Using the same benchmarks, though adjusted to 900 TPC, the researchers estimated that cloud services would quickly become more expensive. An in-house implementation would cost about $400,000 a year to operate, which would increase to $600,000 in 10 years. By contrast, a cloud deployment would start at about $400,000, but would climb to over $1 million a year by the end of the 10-year period.
In addition to comparing cloud prices versus in-house prices, the researchers also examined the costs of hybrid clouds. They found that certain types of hybrid clouds may be more expensive than keeping a system in-house.
Hybrid clouds come in two types, Chul explained. "Vertically partitioned" hybrid clouds are systems in which some of the software (such as application servers) is run in-house, while other programs (such as databases) are run in the cloud. "Horizontally partitioned" clouds are those in which all the software is run in-house, though additional copies could be run in the cloud to meet peak demand.
Overall, vertically partitioned clouds can be expensive to run because of the high data-transfer costs incurred by moving data to and from the cloud, Chul said. Data transfer can account for 30 percent to 70 percent of the costs of cloud hosting. However, a hybrid cloud that does not involve a lot of data transfer could be cost-effective, Chul added. And horizontally partitioned clouds can be a money saver as well, as they could be run only for short periods of time, during peak demand, and not require hardware purchases.
Chul cautioned that the study does not take all costs of a cloud migration into account. Many costs can't be quantified, such as the cost of rewriting applications for the cloud, or the cost of retraining IT help to manage the cloud. As a result, the researchers did not factor these costs into their analysis.
At least one audience member, an engineer from Google, criticized the researchers for downplaying these hidden costs, likening the approach to looking for one's lost car keys underneath the street light, "because that's where it's well-lit."