IBM Corp admits that it does not subscribe to the concept of a single system image for its RS/6000 SP parallel systems a la the Locus Computing Corp definition, but that is precisely because its 900-odd SP customers have indicated that they don’t want a universal paradigm enforced on them, the company says. Most SP users don’t pay the extra for the switch that creates a truly parallel system and simply use the machines as simply VAXcluster master-slave-type clustered RS/6000s.
SP software development manager Dave Turek believes that conventional single system image approaches tend to rob developers of the ability to provide certain types of functionality. In this light, IBM’s work to globalise specific SP resources individually is more consistent with a point product-type strategy. Under the Locus model, most data and processing resources residing on networked nodes (claimed to be about 90% in the case of Locus’s own Transparent Network Computing technology) is presented to users as a single system. Interestingly, IBM once had a partnership with Locus for precisely this technology, but Turek says that after much discussion, IBM decided the pursuit of a Locus-like single system image was ill-advised. So, where Transparent Network Computing creates single process space, Turek says that as far as IBM’s own work is concerned, a unified process space and single Internet Protocol address approach was, never there, never contemplated, never begun and never stopped. He says unified process space models are expensive and difficult to implement. Despite supposed differences in approach, both models appear to share common goals in the areas of system and application management. Systems management from a single image point of view is certainly key to IBM’s effort to capture the local network consolidation (or server consolidation) market with the RS/6000 SP machines, indeed IBM will demonstrate some new single image software (the next PSSP, Parallel System Support Program) at the Supercomputing ’95 show in San Diego. The Parallel System Support Program provides extensions to AIX for installing, operating, and maintaining SP systems from a single point – the control workstation. Its multiple components are integrated into a single software program. IBM will have its other 1995 SP software releases on show in San Diego too, including node isolation – insuring that events that occur and are specific to a single given node do not propagate to any other nodes or elements of the network – plus partitioning and grouping techniques that enable users to achieve very granular system and applications management.
Presently, IBM’s HACMP high-availability clustering software only works across up to eight processor nodes, although that is expected to improve as other availability software is introduced for addressing events that could occur on the system and the responses to them. Turek reveals that the next slew of administration, management and availability software for the SP line will be characterised by ease of use and use of object techniques. He doesn’t expect there to be profound changes to any of the established single image models but says increased scalability will challenge vendors to figure out how to present and administer system information from large-scale installations.