Can a Java Appliance Reduce Power Costs?
Organizations typically depend on scores of servers to handle high processing loads that may occur only once a quarter, or even once a year. But running all those servers means higher costs for power, cooling and space needs.
Those were some of the key issues that Scott Metzger, chief technology officer at TransUnion Interactive, the direct-to-consumer business of Chicago-based TransUnion Corp., the big credit reporting agency, was facing a year ago when he decided it was time to try out some new hardware.
Metzger began testing some Java-based appliance servers from Azul Systems, a Mountain View, Calif., company launched five years ago; it has a proprietary chip set design bundled into its boxes. The chips are specifically designed to optimize applications that are Java-based and run on the Web. The machines also require less energy for cooling than a batch of general-purpose servers with similar computing capacity. The devices are considered more reliable as well, because they operate without a spinning disk inside, Metzger says.
What's more, in a production environment, the Azul machines can handle a much higher processing load for certain types of applications, primarily Java-based systems that run online. The reason is that Azul's Vega 2 chips, containing 48 processors each, can handle processing for high-volume Web applications faster than most general-purpose servers containing x86 chips that have two or four processors.
TransUnion Interactive's Web-based applications include online credit reporting and credit monitoring systems. Through the company's AnnualCreditReport.com site, Metzger says customers can obtain a free copy of their credit report each year. The company also has partnerships with financial firms to provide customers with certain online credit protection services.
TransUnion Interactive deployed Azul servers and was able to reduce the number of general-purpose application servers needed to run its production applications by a factor of six, by offloading the computing load of its customer-facing Web applications. The company has reallocated more than 50 general-purpose servers to other tasks, at the same time eliminating some aging servers with high-memory configurations. "This server reduction has had measurable impact on our rack space, power and cooling requirements to support these applications," Metzger adds.
By getting rid of so many servers, TransUnion Interactive saves on power and cooling costs for numerous pieces of hardware it would otherwise have to run. Rack space has been reduced by about two-thirds.
"Today we are using Azul for the majority of our online applications," Metzger says. "Our Web-based applications really benefit from the Azul hardware because of their highly multi-threaded processing capabilities."
In a Web application, each concurrent user has his own execution thread, which can really bog down most general-purpose servers as hundreds of threads are generated, Metzger explains.
But the Azul servers are multi-threaded, affording much faster processing speeds as the application runs across multiple processors. "With Azul, the throughput you can achieve is much greater," he says. "We have 10 applications on that architecture, and if one application spikes demand, we can reallocate memory and processing services to that application.
"It allows us to be much more nimble in allocating processing resources to whatever application needs them to maintain service levels. From an operating system perspective and in terms of general hardware maintenance, it's very effective for us to be able to scale the data center at the least cost, while essentially providing a much larger sandbox for our applications."
The Azul machines use network-attached processing, providing centralized computing and memory capacity to Java-based applications the same way network-attached storage provides centralized storage capacity. Although the applications launch from traditional servers, they connect via a gigabit Ethernet subnet to the Azul machines, which provide up to 768 processor cores and 320 gigabytes of memory. The Java application workloads are offloaded from the traditional servers to the Azul Virtual Machine, which is part of each appliance. Once offloaded, the application gains access to larger memory and as many processor cores as it needs. The access is controlled by the Azul Compute Pool Manager, which the user sets according to the desired allocation of computing resources.
In fact, the Azul servers tend to do their best on Web-based Java applications.
"The Azul systems are optimized for Java code," remarks Gordon Haff, principal I.T. adviser at Illuminata, a research firm in Nashua, N.H. "Running that kind of a workload, they can save companies significant power and space. This is especially useful for large financial institutions that need to scale for very large workloads while ensuring that they maintain consistent response times."
Making changes to the applications running on fewer machines is faster, too. "For example, when we have to initiate a rules change due to Daylight Savings Time and we have to apply those rules to both our production and non-production servers, it's a much smaller task to do it on six devices versus 60 servers," Metzger says.
Lighter on The Juice
Regarding power consumption, Metzger has found that in a comparison of 4 rack-units of Azul machines with 4 rack-units of general-purpose servers, Azul draws about 550 fewer watts of power846 wattsrepresenting about a 40% savings.
A single Azul model 3220 requires about 3,500 BTUs per hour to cool; a total of seven dual-socket Intel x86 servers with the same computing capacity require a total of 7,000 BTUs per hour, according to an Azul spokesman.
TransUnion Interactive also has put an additional Azul appliance to work in a non-production environment supporting the efforts of software development teams.
"Now we have one appliance that supports up to 10 development teams," Metzger says. "Using general-purpose hardware, we would have needed 25 to 30 servers to handle that workload."
Going forward, Metzger plans to continue to consolidate TransUnion Interactive's server hardware in an effort to further drive down cost per rack unit, including power and cooling costs. "I expect to cut our fully loaded rack-unit costs in half," he projects.
On the downside, an Azul server will set a company's information-technology budget back about as much as a midrange BMW. Prices start at $49,995 for a 96-core, 48-gigabyte appliance and go up from there.
But as Metzger and TransUnion Interactive found, each Azul machine can replace a good-sized bundle of general-purpose servers.