IBM supercomputer perspectives

ASCI White is the fastest SP and the fastest computer in history. It is installed for the U.S. Department of Energy (DOE) and will deliver 12.3 TeraFlop/s peak performance, more than the 10.2 TeraFlop/s of the contract. Michael Teyssedre compared 1 second ASCI White with 10 million years of hand calculator. It is equivalent in power to next four supercomputers o the Top500 list and can process an online transaction for everybody on earth in less than one minute. The floor space is equivalent to three tennins courts and the weight of 17 full size elephants. The 8192 copper microprocessors need 2000 miles of copper wiring, comparable to the distance of Paris to Moscow. The computer is made of available parts, RS/6000 SP high nodes, SPS 2 switch and is committed on site upgrade.

IBM has sold 10 000 SP systems with 100 000 nodes. The receipt for the success was, following Michael Teyssedre, the Power processors, the Switch and the linear scalability to 512 nodes of SMPs. An other aspect is the manageability using PSSP (Parallel System Support Program). As demonstrated by DeTeCSM, all can be managed from one operator console. An other important aspect is the binary compatibility of programs. IBM ran the same code on an RS/6000 320 from 1990 and on a new RS/6000 270. The only difference was the performance, as the new system ran abot 100 to 150 times faster.

The Future of SP and the Power4 Chip

The future is built on SP foundations, the Power4, the switch technology and PSSP becomes pervasive. The interconnect is brought right down to the chip. The Power4 - the Gigaprocessor - uses copper, SOI (Silicon on Insulator) and is produced in 0.18 micron. Two CPUs are on a chip, > 1 GHz. There is a chip to chip interconnect of > 500 MHz and > 35 GBytes/s. The transfer from the L2 cache is > 100 GBytes/s, to the L3 cache > 10 GBytes/s. The processor is NUMA enabled. An 8-way SMP is in the palm of ones' hand. The 32-way SMP is of 30x30 cm size. Michael Teyssedre reported that the power is on on the test chip since 4Q/98, two 32-way systems run in the IBM labs. He expects the delivery in October 2001.

On the road to Power4 each server will be part of the SP family, Hxx, Mxx and Sxx enterprise servers and the SP systems are integrated, the best of both worlds. Thus IBM will be faster in time to market and more flexible to fit the customers' requests. The adds in the future are the enterprise server clusters, more powerful building blocks, more configuration flexibility, cluster software enhancements, faster interconnect, LPAR (partitioning of the parallel system), NUMA, clusters of SMP and of NUMA.

Unix Server Evolution - "Regatta"

Michael Teyssedre described the most important factors of this evolution. First there is the granular scaling. This means that the performance is balanced between processors, memory, I/O and switch interconnect. Then it scales from one to thousands of processors and RS/6000 servers can be attached and clustered. The management of the system is flexible with a workload manager, physical and logical partitioning and the most important tools are available. The system is built out of common building blocks, Power4 processors, 8- to 32-way SMPs, 64 to 256 GB memory and the high speed interconnect. It is possible to connect machines directly by a NUMA connection - the Sequent technology. The other possibility is the connection via the Federation Switch, which is realised for the ASCI program. This switch has 100 times better throughput than the actual switch, about 32 GByte/s.

The Federation Switch allows an interesting approach. On the one hand one can connect Power4 processors and SMP, running AIX and Linux. The most surprising thing, Michael Teyssedre presented is the attachment of Intel processors. It will be possible to connect McKinley and Foster - IA64 - processors. These are low cost building blocks running Windows 2000. It will be possible to operate this machine from one console and manage the multiple operating systems. He mentioned that this tool is still running in the IBM labs.