SGI has emerged from a major reorganization with a renewed focus on its core scientific and technical markets, including the life sciences, according to Deepak Thakkar, the firm’s bioscience segment marketing manager.
The company, which filed for bankruptcy protection last May and emerged from Chapter 11 status in October, has a new management team that “has charged us with developing solutions for specific industry segments so that we can address the needs of specific groups of customers,” Thakkar said.
The backbone of this industry-focused approach is a hybrid high-performance computing platform that blends three different computing architectures — x86 clusters, shared memory systems, and the company’s RASC field-programmable gate arrays — into a single, integrated system.
While the company expects this approach to appeal to a wide range of high-performance computing markets, Thakkar said that it is particularly well suited to the bioscience sector, where research groups often rely on a wide range of software packages that are all optimized for different computing architectures.
Sequence alignment programs like Blast and Smith-Waterman, for example, run well on clusters, and even better on FPGAs, while molecular modeling and virtual screening applications require a shared memory architecture.
Thakkar said that smaller biotechs and research groups on a limited budget must currently choose one type of system over the other. “No matter what they buy, they will find that some of the software they use will run suboptimally on any system that they buy,” he said. “By investing in a single system, they’re in fact making their work a lot more difficult because now they’ve pretty much got a suboptimal system running throughout the organization.”
Thakkar said that the single-rack system “appears as a single image for the system administrator,” who can then allocate space on the shared memory system for those codes that require a large amount of memory, and on the cluster for codes that require parallel processing. Larger, more departmentalized organizations can still benefit from the integrated system by partitioning a single shared system, he said.
Thakkar said that a typical biosciences implementation under this hybrid approach would be priced “competitively” with single-architecture HPC platforms from competitors like IBM, HP, and Sun.
As an example, he said that a 16-processor cluster system with 16 GB of memory and almost 4 TB of storage would start at under $30,000. This system could be scaled up “seamlessly” to include additional cluster compute components, shared memory, FPGAs, or additional storage, he said.
SGI expects the system’s flexibility to be its key advantage in the marketplace. “This is a unique architecture,” Thakkar said. “No other organization — not HP, not Sun, not IBM — has this kind of tri-partite system, where in a single rack you have the capability to do all three kinds of computing that is required across the organization.”
Thakkar stressed that every installation is customized to ensure the right mix of architectures for individual research environments. “There is no single solution that is going to fit everyone,” he said. “It’s a sliding scale, and we have the ability to increase or decrease the amount of cluster, shared memory, or RASC. Some customers can get away with just the clusters and the FPGAs, or they may need a little bit of everything,” he said.
“It’s a sliding scale, and we have the ability to increase or decrease the amount of cluster, shared memory, or RASC. Some customers can get away with just the clusters and the FPGAs, or they may need a little bit of everything.”
Thakkar said that SGI has partnered with InforSense, the Scitegic subsidiary of Accelrys, Teranode, the BioTeam, and the open source Taverna project to ensure that a range of life science workflow packages are fully integrated with SGI’s servers to serve as a front end for end users.
In addition, the company has an ongoing collaboration with Mitrionics, which has already released one FPGA-accelerated implementation of Blast that runs on the SGI RASC architecture, and just announced last week that it has started developing an accelerated version of BlastP [BioInform 02-23-07].
Other accelerated bioinformatics applications are also coming online for RASC. Walid Najjar at the University of California, Riverside, is developing a compiler that translates C code into VHDL, the language used to program FPGAs. One of the first test cases his group is working on is an implementation of the Smith-Waterman algorithm that has run up to 4,000 times faster on RASC than a typical Xeon or Opteron in preliminary tests.
Najjar told BioInform that his group is finalizing the implementation and expects to release it for use on RASC “soon.”
Najjar’s team recently purchased an Altix XE240 system with a RASC RC100 blade, which is the “first major machine” that the researchers have ported their code to, he said. His team has been doing its development work on Xilinx FPGA boards and other systems that Najjar described as “small toys compared to RASC.”
SGI’s Thakkar said that UC Riverside was one early customer for the hybrid architecture, which the company just began releasing formally at the beginning of the year. Other life science customers for the hybrid system include the University of Arizona and the National Cancer Institute, as well as several undisclosed commercial firms, he said.