FRAMINGHAM (10/21/2003) - Pharmaceutical giant Novartis AG is undertaking an ambitious five-year plan to harness the unused processing power of every desktop computer in its organization to speed research efforts. When it's finished, more than 65,000 PCs will be connected into a giant grid, but already a pilot project involving 2,700 PCs is proving its worth.
This summer, Novartis researchers reported that they used less than a week's worth of the small grid's processing time to identify a previously unknown CK2 inhibitor in the company's compound library. Not bad for a project begun modestly last spring with only 50 PCs knitted together. The 50 grew quickly to 2,700.
Grid computing is the centerpiece of Novartis' long-term effort to expand R&D computing resources. Eventually, the company expects to capture the "spare" processing power -- estimated to be about 90 percent of a desktop computer's capacity -- of all of its PCs worldwide.
"We did not have articulated performance goals," says Manuel Peitsch, head of informatics and knowledge management in Novartis' research division. "It was an experiment and we just expected more things to be done because of the grid." The early speculation was that grid computing could help Novartis identify up to 10 times more targets than before.
Novartis researchers describe in the June 19, 2003 issue of Journal of Medicinal Chemistry ("Discovery of a potent and selective protein kinase CK2 inhibitor by high-throughput docking," Vangrevelinghe, E. et al. J Med Chem 46, 2656-2662) how the grid was used to screen the corporate library of compounds and found an unknown CK2 inhibitor. "So less than one year, and already we have our first (finding)," Peitsch says.
Currently, Novartis is running applications such as BLAST and high-throughput docking and screening simulations on the grid. Custom routines that do gene comparisons are also being ported to the system. In the future, Novartis even intends to use the grid for knowledge extraction and text mining.
This latter application represents an innovation. Typically, grids are used to do intense number crunching, not text searching. According to Peitsch, "We have lots of text-based information. You could send one document (in a large collection of documents) to each CPU and look for text."
Further down the line, Novartis hopes to leverage the grid in clinical trials. Peitsch notes there are lots of statistics in clinical trials. The grid might be useful in simulating trial results to look for what he calls the boundary values of a particular trial to define or determine the ideal patient population and drug dosage.
Novartis is working on several fronts to make the corporatewide grid happen."The first stage is to create a grid with desktop hardware in each of our geographies," Peitsch says. He notes that they've restricted the deployment to desktops and do not include laptops. "The main reason is that laptops are off the network too long," he says.
Novartis is using Grid MP Enterprise software from United Devices to manage and run the grid.
Currently, the grid delivers about 5 teraFLOPs (trillion floating-point operations per second) of processing power. The company spent about US$400,000 for grid software licenses, but estimates it saved about $2 million to get that amount of processing power.
At the same time, Novartis is upgrading its 65,000 PCs to a standard hardware configuration and software desktop. The platform of choice is an HP Pentium 4-based PC running Windows XP. Every new platform must conform, so as to minimize system management costs and ease application development.
"Standardization is key in the deployment of a large grid," Peitsch says. "If it works on 10 desktops, it will work on 100 or 1,000." Another benefit includes reduced unit pricing because of the higher purchase volumes.
Five years ago, Novartis standardized on Windows NT, and Peitsch estimates that move saved $120 million in total cost of ownership over a three-year period. The company expects to save 1,000 Swiss francs per year for the PCs that were standardized on NT, and 2,500 Swiss francs per year on legacy PCs that were not standardized but will be over time. This could result in another $200-million savings over three years, he says.
For instance, by having common desktop hardware and software worldwide, Novartis expects to reduce the number of helpdesks for desktop support from "tens" to "two or three" for its entire enterprise. Peitsch concedes that rolling out a grid to 65,000 PCs is likely to produce a few unpleasant surprises.
At present, every PC in the Basel, Switzerland, office is connected to the grid, and the grid is being rolled out in the company's new facilities in Cambridge, Mass. "As PCs are delivered to employees, they are automatically part of the grid," Peitsch says. And the grid in the U.K. office is "on the way."
One concern with deploying such a large grid is its potential adverse impact on corporate network performance. So far, this hasn't been a problem. "The bandwidth load of running this grid is fairly marginal," Peitsch says, though he warns that application design is critical. "Don't move around a lot of data and have lots of data crunching on each machine."
Clearly, grids can't handle all computing tasks, Peitsch agrees, but "(they've had) a tremendous impact. People in research are now doing things they did not think were possible before."
Every little bit helps
Number of PCs in the grid today: 2,700 and growing
Types of applications supported: Common bioinformatics routines like BLAST and docking simulations
Processing power: 5 teraFLOPs (5 trillion floating-point operations per second)
Payoff thus far: Docking simulations found an unknown CK2 inhibitor in corporate library of compounds
Applications being explored: Knowledge extraction and text mining performed on large collections of documents
Number of desktop PCs being replaced in ongoing upgrade: 65,000 corporatewide
Standard desktop environment once upgrade is complete: Hewlett-Packard Co. Pentium 4 PC running Microsoft Corp.'s Windows XP