India agency offers build-to-order supercomputer

The Indian government-run Center for Development of Advanced Computing (C-DAC) has designed a parallel-processing 1 TFLOP (trillion floating point operations per second) supercomputer, scalable up to 16 TFLOPs and available on a build-to-order basis.

C-DAC was set up in 1988 with the objective of designing a supercomputer, after India's bid to purchase a supercomputer from the U.S. for weather forecasting, fell foul of U.S. restrictions on exports of high performance computers to India.

The first supercomputer from C-DAC, the PARAM (for PARAllel Machine) 8000 was introduced in 1991 with a rating of 1 GFLOP (billion floating point operations per second). Supercomputers of ever-increasing processing power followed from C-DAC.

These supercomputers are typically built around standard components, except for a few components like the communications co-processor, which was designed by C-DAC and fabricated abroad. While a 100 GFLOP PARAM is designed around UltraSPARC II processors from Santa Clara, California-based Sun Microsystems Inc., the new 1 TFLOP supercomputer, called the PARAM Padma, uses the Power4 processor from IBM Corp. in Armonk, New York, in a symmetric multiprocessor configuration. "The decision to shift from UltraSPARC-II to Power4 was on techno-economic grounds," said Raj Kumar Arora, executive director of C-DAC, based in Pune, India.

India is included in the Tier 3 of the U.S. HPC Export Control Policy of the U.S. Department of Commerce (DoC). Although the U.S. government relaxed in March this year the upper performance limit of computers that could be exported to India from 85,000 MTOPS (millions of theoretical operations per second) to 190,000 MTOPS, imports of supercomputers comparable to the new 1 TFLOP computer designed by C-DAC are still restricted, according to Arora. "The performance of the PARAM Padma in terms of MTOPS is in the vicinity of 500,000 MTOPS," Arora added.

None of the PARAM supercomputers installed in India so far are used in defense organizations, according to Arora. "We have maintained throughout that our research is for civilian applications, and not for defense and nuclear applications," said Arora. "Some defense organizations in India have their own supercomputer projects. We see the PARAM project as helping build our self-reliance, and also to help establish India's hardware design capability."

The Indian government is investing about USUS$30 million over the next five years in its supercomputer program and on setting up a grid computing network. "The U.S. export controls were at the top of our mind when we started the supercomputer program in India," said Rajeeva Ratna Shah, secretary to the Indian government's department of information technology. "We have evolved since then into designing cost-effective and open systems supercomputers that have a market both in India and abroad, for a variety of applications both commercial and scientific."

Of the 52 PARAM supercomputers installed so far, only seven have been installed outside India. Two of them were installed in universities in Canada and Germany, a third at a financial modeling and simulation firm in Singapore, while the rest were sold to Russia.

The low cost of the PARAM Padma, which is priced at about US$5 million, and the applications C-DAC plans to offer on the platform, will place it at an advantage in the international market, according to Arora. "We will develop applications depending on customer interest," Arora added. Currently, applications in bio-informatics, computational structural mechanics, computational atmospheric science, seismic data processing, computational chemistry, evolutionary computing, and computational fluid dynamics are available on the supercomputer.

The PARAM Padma runs on both the Linux operating system and AIX, IBM's version of Unix. The AIX version supports 62 nodes with four processors each, while the Linux version supports only eight nodes, as Linux is not able to scale over more nodes, according to Arora.

The nodes on the PARAM Padma are connected through a primary 2.5G-bps (gigabits-per-second) full-duplex system area network (SAN) designed by C-DAC, and a Gigabit Ethernet backup network. "We had to design a switch from scratch for inter-node data traffic, rather than use a Gigabit Ethernet switch, because the performance of the system largely depends on the performance of the switch," Arora said.

The storage system provides a primary storage of 5 terabytes scalable to 22T bytes. The storage architecture uses Fibre Channel-Arbitrated Loop (FC-AL) based technology for interconnecting storage subsystems like parallel file servers, NAS (network attached storage) servers, metadata servers, RAID (redundant array of independent disks) storage arrays, and automated tape libraries, with an I/O performance of up to 2G bytes-per-second.

C-DAC is also planning to set up a national computing grid that would have as its participants key academic and research laboratories in India. The resources from the grid, which will have 10 TFLOPs of compute power and 1-petabyte (1 million gigabytes) storage, will also be leased out to the corporate sector. C-DAC is codeveloping grid computing technology in collaboration with IBM. The partnership with IBM also includes codevelopment of speech technologies and bio-computing, according to Arora.