Updated IBM software to cluster more units, over distance
- By Sari Kalin
- Sep 01, 1996
IBM Corp. is upgrading clustering software for its RS/6000s enabling users to cluster more machines and spread out machine availability over long distances an IBM official said.One initiative code-named Phoenix will expand the number of RS/6000s that can be clustered from eight to 32 and next year 128 according to David Turek director of Scalable PowerParallel (SP) Systems in IBM's RS/6000 division.
Phoenix to be announced in October will ship as a feature on the SP later this year Turek said. Early next year IBM's High Availability Clustering Multiproc-essing (HACMP) software for SP systems and general-purpose RS/6000s will be upgraded to include Phoenix Turek said. HACMP will retain the same interface he said.
IBM will be employing RS/6000 processors in an SP2 system the company is building for the Energy Department's Lawrence Livermore National Laboratory. The company will deliver the machine under a $93 million supercomputer contract the lab awarded in July [FCW July 29].
Phoenix will be able to handle operating system middleware and application failures not just hardware failures Turek said. IBM also is working with OS middleware and application software vendors to develop standardized application program interfaces for Phoenix Turek said.
"Everybody thinks of [availability] in terms of hardware " he said. "We've built an infrastructure capable of handling arbitrarily defined software and hardware events."
Also in October IBM plans to introduce a HACMP-derived product that will allow geographic availability clustering Turek said. For example if a machine fails in Manhattan a recovery script could transfer work to a machine in Boston. IBM already is testing systems over a 1 500-mile range Turek said. Geographic availability will be limited to eight-machine clusters initially he said.
In another announcement IBM said it plans to open up the switch that links its parallel processing supercomputers so that customers can add auxiliary devices that improve performance.
The first product to use the opened switch will be a high-performance gateway node - essentially a souped-up router that can bring large amounts of data from an outside database to the RS/6000 SP systems according to IBM.
Opening up the SP switch to devices outside the SP box makes the SP more flexible IBM said. In this "logical SP" architecture outside devices appear no different to the user - or the system administrator or the SP's management software - than resources inside.
"It's broadening the reach of the technology customers are able to bring to bear on their business problems " Turek said. "Ultimately it's improved performance."
For example the gateway node will be a more cost-effective way of bringing data to the SP than the current architecture in which input/output connections are made to individual nodes Turek said.
IBM developed the gateway node with a router vendor that Turek declined to name. The node will be announced formally in November at Supercomputing 96 and will ship in the middle of next year he said.Future additions to the "logical SP" architecture could include other kinds of I/O devices as well as ultralarge symmetric multiprocessing systems Turek said.
In the future IBM also plans to focus its SP sales pitch more around solutions Turek said. "You'll hear less and less talk about clusters and more about Web servers that have a cluster incarnation " Turek said.