The Phoenix technologies which IBM Corp is creating for use with its parallel SP and RS/6000 SMP servers will spawn a slew of new high-end software offerings in addition to the 128-way clustering already identified (CI No 2,985). To make the SP series more attractive for file serving IBM will add a new parallel file […]
The Phoenix technologies which IBM Corp is creating for use with its parallel SP and RS/6000 SMP servers will spawn a slew of new high-end software offerings in addition to the 128-way clustering already identified (CI No 2,985). To make the SP series more attractive for file serving IBM will add a new parallel file system option for AIX next year, plus an interface reportedly known as Perspectives that will allow administrators to drag and drop objects between nodes. In adding support for 32-way SMP clustering, Notes and Web servers and the existing High Availability Cluster Multiprocessing APIs, Big Blue hopes that some of the SPs success will rub off on the rest of its Unix server business. For example it’s investigating how SP nodes can be opened up to accommodate any RS/6000 model; how the physical distance between nodes can be increased and how to allow external nodes attached to SP parallel switches to be viewed and managed from within an SP cluster using Phoenix; plus how to get more data and user traffic flowing into SPs via multiple network connections. To accommodate the latter, which would allow users to create faster data warehousing or data mining environments, IBM will sell a router developed in conjunction with one of the major players, thought to be Cisco Systems Inc, called the High Performance Gateway Node. Attaching the network concentrator directly to the SP switch as is planned will allow multiple data streams to flow into an SP system without burning individual RS/6000 processors and AIX licenses just to handle external I/O networking tasks. The router won’t be integrated into SP’s cabinetry but will be managed by existing management applications as if it were. Putting together a business plan to sell and support the third party device means ships aren’t likely to begin for at least six months and maybe up to a year.
What IBM refers to as other out-of-box SP options will eventually allow customers to attach any RS/6000 to the SP switch and have it act as an integrated node. That might extend to supporting non-AIX and indeed non-IBM environments too, though the company admits that depends in part on the extent of management software capabilities and development of SP switch and processor technology. Already planned are next-generation versions of the SP switch which will use the future RS/6000/AIX Mezzanine memory controller to attach memory directly to the switch rather through I/O connectors, overcoming MCA and PCI bandwidth limitations. Phoenix and the new switch will lead into other distributed memory and crossbar architectures; we’ve learnt of one called S-COMA. On the solutions side, the SP group is getting closer to OLTP provider BEA Systems Inc on a variety of configurations of the Tuxedo transaction processing monitor, including a Tuxedo with Oracle strategy.
IBM says it had 1,515 RS/6000 SP parallel systems installed at the beginning of August with a total of 16,000 nodes, 1,000 or more of them commercial systems and 500-plus technical units. There are more technical nodes installed – that’s where the really CPU-hungry users are found, including Cornell University’s 512-processor beast – though IBM claims there is one commercial user at nearly 400 processors and many others at the 50-to-60 CPU level. By the end of 1996 IBM hopes to have doubled its 1995 SP installed base, which stood at 1,023, which means it has got 531 to go. While critics say many SPs are deployed primarily as LANs-in-cans – shared memory rather than shared nothing devices – SP supremo Dave Turek claims there are only a handful of dozens of users that don’t use the interconnect, now at 113Mbps and 37 microsecond latency, all of them commercial installations. In these cases, SPs are generally being used to consolidate large-scale server operations, he says.