Linux-based grid enables Hewitt Associates to get more out of its mainframe
The mainframe is alive and well at the global human resources consulting and outsourcing firm Hewitt Associates, but it’s getting quite a bit of help these days from a Linux-based grid computing platform.
For one application, the grid reduced total computing costs by about 90%, according to Daniel Kaberon, director of computer resource management for Hewitt, based in Lincolnshire, Ill. For another, it reduced the time required to run jobs by as much as 95%.
Kaberon spoke about his experiences with the DataSynapse GridServer platform at the recent Network World IT Roadmap Conference & Expo in Chicago, as well as in a follow-up interview (available here as a Webcast).
Some 2,300 companies use Hewitt Associates consulting services while another 340 companies rely on the firm to administer human resources, healthcare, payroll and retirement programs to millions of their employees and retirees. Hewitt has 24,000 employees in 35 countries.
Hewitt likes its IBM zSeries mainframe for a variety of age-old reasons, Kaberon said, including security, scalability, auditing and high availability, to name a few. “The great joy of having mainframe applications is they can have a lot of compute power, storage, database and transaction processing all in the same place,” he said. “It gives you a big pool of resources to do things with a tremendous amount of scale and availability.“
The downside is cost. “For certain applications that don’t exploit those qualities of service, using that compute power is a very expensive way to get things done,“ Kaberon said.
DataSynapse GridServer software lets Hewitt selectively offload some application processing chores from the mainframe to relatively low-cost Linux-based blade servers, then ship the results back to the mainframe. In that fashion, Hewitt maintains all the benefits of the mainframe while dramatically lowering its costs.
Pension app test bed
Hewitt began dabbling with grid computing in 2003, starting with its Pension Calculation Engine application. Hewitt provides pension administration services for many Fortune 500 companies. The pension application enables employees of those firms to calculate their expected benefit payments.
“[Employees] want to know how much would they get per month if they were to retire when they turn 60, or later this afternoon after a terrible bout with their boss,“ Kaberon jokes. The requests come in over the Web and the application collects information about their plan, their hire date and so on, then crunches the numbers to come up with an answer. “That takes a lot of calculation power to do,“ Kaberon said, placing a large CPU cost on the mainframe. Should the individual want to change some parameters — say, if they retire at 65 instead of 60 — the process starts all over.
Compounding the problem for Hewitt was that the load on the application is hard to predict, as good or bad news at any given client company often triggers a barrage of hits to the pension calculator from that company’s employees.
To address the problem, Hewitt — working with IBM and DataSynapse — recompiled the application, which is written in Smalltalk, to run on the Linux-based blade server grid. The mainframe collects the personal and plan data required from its database, then uses a Simple Object Access Protocol connection to send the data to the Linux grid, where the calculations are performed. Once the calculations are complete, the results shoot back up to the mainframe, from where they are dispersed to the requester via the Web.
The application lent itself to the grid configuration because, other than gathering the required data from the mainframe database, it could perform its calculations independently. That meant few changes were required to the application itself. From project kickoff to production, the implementation took eight months.
“And we can scale very easily because of the way GridServer product works,” Kaberon said. “We can just add more horsepower and it continues to spread the work across without having to move around the data.”
The setup offloaded about 1,800 MIPS from the mainframe and reduced costs for the pension calculation application by 90% vs. using the zSeries mainframe alone, factoring in the costs of the Linux grid and DataSynapse software, Kaberon said.
Perhaps more importantly, it established the credibility of grid computing at Hewitt, leading to more such projects with similarly impressive results.
Powerful for print
The other project that Kaberon explained to the IT Roadmap audience involved a print application from Sefas Innovation. Used to print customer statements and the like, the application generates enormous amounts of postscript and PDF output. Additionally, the application was single-threaded, meaning it could take advantage of only one CPU at a time.
To prepare the application for the grid environment, Hewitt developed a splitter program that breaks up the input data into 2.5MB chunks and places into a Network File System (NFS) input directory on the mainframe. The parcels are then distributed to the grid for processing and the output is written to an NFS results directory, also on the mainframe. There the results are combined back into a single final object.
“Using the distributed DataSynapse environment we were able to take that [application] and do an enormous amount of stuff at the same time,” Kaberon said. One job that required about 300 hours to run on the mainframe was completed on the grid in only 14 hours, a nearly 95% improvement.
Choose your spots
Not all applications will see those kinds of results, however. Applications that lend themselves to grid computing are those that use a lot of compute power but little I/O, whether to files or databases. “There’s a certain amount of overhead to ship a request back and forth, on and off a platform,” Kaberon said. If an application has to continually shuttle requests back and forth, the grid won’t make sense. “If you can send a request intact without out having to do more I/O, you’re more likely to have a significant payback.”
For Hewitt, the grid has let it free up capacity on its mainframe, which is worth “a significant amount of money,” Kaberon said. “It created a lot more headroom for other applications and the growth that our businesses experience, so we could handle that with a small increase in cost. The cost of the grid implementation is very small compared to the capacity that we freed up.”
Hewitt continues to look for applications that are a fit for the grid, such as applications that don’t necessarily require all the qualities of service that the mainframe offers, with data that naturally lives on the mainframe but that can be moved elsewhere as necessary.
“We also need to have cooperative partners,” Kaberon said. “If we’re working with applications written by third parties and they’re not great believers in [grid computing], then it’s not sensible to do it.”
Kaberon does have advice for other would-be grid implementers. “Keep your eyes open,” he said. “Don’t get stuck on the idea that the place where data is born and living means that’s necessarily the best place to process it.”
Desmond is events editor for Network World and president of PDEdit, an IT publishing company in Southborough, Mass. Reach him at paul@pdedit.com.
