Last month Grid Computing Now! hosted an informative webcast seminar entitled “The Business Case for Next Generation IT Architecture.” On hand to share their experience and expertise were Steve Wallage, Research Director for the 451 Group, and Shahid Mohammed, EMEA Database Team Manager for Marsh.
Wallage starts out by addressing the confusion over Grid terminology. He cites a 451 Group poll, in which 70 percent of respondents felt that there was a better term than “Grid computing” to call their distributed computing architecture, preferring the terms high performance computing (23 percent), virtualization (23 percent), clustering (19 percent), utility computing (19 percent), and service-oriented architecture (13 percent).
Another survey asked over 250 early adopters what made them think about using a Grid computing strategy and employing a Grid architecture. “Improved performance” was followed closely by “saving money” as key business drivers. For the oil industry, saving money in terms of oil exploration is important. It can cost up to 100 million dollars just do dig a hole. Grid computing allowed the industry to improve their seismic modeling from 3D to 4D, which led to a more effective oil exploration strategy.
But the reasons went beyond saving money, to the realm of making money. Grid can help companies get a leg up on the competition. Wallage gives the example of a company that used Grid computing to compute insurance costs and risk premium on an hourly basis during Hurricane Katrina. This gave them an advantage over their competitors who could not provide premiums and risk management until after the hurricane ended.
Wallage discusses one company that tried to determine the quantifiable and unquantifiable benefits to deploying a grid. Tangible benefits include hardware expense and better software license utilization. Intangible benefits include better collaboration, better use of resources, and the ability to do new and better things.
Wallage also addresses the challenges of deploying Grid. Software licensing is a big concern for a lot of companies. In some industries, the typical licensing model is per-node or per-computer. In EDA (electronic design automation) one license can cost up to a million dollars. Multiply that by a 100-node grid and you get an idea how costly this can be. It can defeat the financial incentive of deploying a grid in the first place. Users try to find ways around this, such as running software only on a select number of devices, buying enterprise license agreements, or negotiating with the vendor for a different model.
Cultural challenges cover a wide range of issues. One of the problems of going to a shared-ownership model is convincing all parts of the company that it's in their best interest. An IT department is used to running their own IT and can balk at a shared architecture model.
Wallage reports that the challenges grow as an organization goes beyond the compute grid: “Typically the compute grid is used for those fairly easy low-hanging applications that run well on a grid.” These are typically applications that involve Monte Carlo simulations, drug discovery, and seismic processing, which tend to be easily componentized, and easily parallelizable. According to Wallage, the obstacles mount as one goes beyond these areas, towards more data-driven applications.
Metering, billing and chargeback offer another challenge. “As you develop a central architecture, how do you accurately and fairly charge different parts of the company for the IT resources they are using?” asks Wallage. Raw compute power is easier to manage, but networking, storage, and software license usage are not so simple. Wallage explains that's why some companies stick to simple things, such as compute hours, to work out how to meter different parts of the company.
He envisions a five level model of Grid. Level one entails trials and proof of concepts. Level two involves running a single application, a “no-brainer” application like the Monte Carlo simulation. Level three includes siloed grids (single grids, or grids in multiple departments that aren't linked). Level four involves linked grids — running multiple applications across multiple areas. An example of this is an HPC Grid, purely for HPC applications. Level five is the final level. It constitutes an internal shared utility and a platform for shared services. Investment banks are often deploying Grid at this level.
“Level five is where the real potential advantages lay and also where it clearly becomes more than just Grid. Virtualization, SOA and other technologies all start to come together,” reports Wallage.
Helping companies achieve level five are virtualization, utility models, server consolidation, open source, SOA, and enfranchising existing applications.
Wallage explains how Grid and virtualization are related saying, “It's not a question of Grid versus virtualization, it's really how the two combine.” He explains that Grid is a way to aggregate resources using virtualization techniques and scheduling. Virtualization and SOA are the drivers of next-generation Grid development. Some vendors are dropping the word Grid from their descriptions, but the term is still relevant since it becomes part of the infrastructure, and it is used as a means to an end.
Wallage says the 451 Group acknowledges three basic types of virtualization: partitioning, emulation, and aggregation. Partitioning splits one resource and makes it appear as many. Emulation makes one resource appear as something else. Aggregation (which is most like Grid) gathers many resources and makes them appear as one.
There are other levels, such as application virtualization, which sits on top of the stack. Many so-called Grid vendors, for example DataSynapse and Platform Computing, are increasingly focused on application virtualization. They no longer see themselves as either Grid or traditional HPC providers. Rather, they provide an application fabric layer. They enable companies and users to put one application across an infrastructure and for the application not to be aware that it's running on a Grid-type infrastructure. Wallage thinks this is the Holy Grail for the marketplace.
While Wallage gives a broad overview on the current state of Grid, Shahid Mohammed, the EMEA Database Team Manager for Marsh, deals with the practical realities of Grid implementation for an eCommerce company. In 2003 his company had a typical data center with a 3-tier set up with application servers, database servers, and storage. Typical cycles of upgrades for applications were 3-6 months, which meant throwing away application servers and database servers and storage and upgrading to the next model. The applications were stovepipes: a certain set of application servers were dedicated to online transaction processing and another set of database servers were dedicated to batch processing, which would happen overnight. And Mohammed notes that the setup “didn't really bare any relation to what exactly the tasks that needed doing were, more to what the budget of the project was.”
Mohammed's company needed to be able to cope with a yearly growth rate that was 40+ percent year-on-year, respond better to project requirements, consolidate databases, and get a better return on investment. Their return on investment was 19 percent and they did not think they were getting good value for money. Mohammed also needed to improve manageability, saying “We wanted to reduce our technology stack to make it easier for us to manage what we had.”
According to him there were two possible solutions: “Buy one mega-high-end server and throw as much as we could onto that, or buy smaller servers and create a grid.” They weighed their options and, taking cost into consideration, went with option two.
They did the storage first and then the applications and the database tier at the same time. For the database tier Mohammed implemented an Oracle 10 rack, which allowed the company to create a database pool with many machines.
Mohammed reports that the final Grid solution was achieved in 2005 and, most strikingly, provided a 70 percent return on investment. It allowed for dynamic resource allocation, allowing the batch applications to use more resources overnight and the online transactions to use more during the day. They realized a lower cost for capacity, and the projects now pay for capacity, not hardware. Easier maintenance was achieved by trimming off the different storage units and database servers and replacing them with a common architecture.
A question and answer period rounds out the informational webcast, covering such relevant topics as cultural challenges and training, data center energy issues and security concerns.
To see the webcast visit the Grid Computing Now! website at www.gridcomputingnow.org or access the webcast at http://brighttalk.com/comm/gridcomputingnow/325a2ebf71-2473-335-2284.