Early next year Sun Microsystems Inc will outline new technologies – dubbed Serengheti – that it will offer for harnessing the processing power and memory of multiple systems by linking them together so they behave as a single virtual processing node. Call it ccNUMA by another name, but Sun doesn’t want its new 3D architecture, the basis of Serengheti, sullied by association with the current crop of cache-coherent distributed shared memory systems offered by the likes of Sequent Computer Systems Inc, Data General Corp or Silicon Graphics Inc. It thinks ccNUMA is fundamentally flawed because it introduces a bug into the shared memory ‘uniform’ model of SMP symmetric multiprocessing, namely a latency, or time penalty for accessing a remote memory store. It claims users and ISVs must undertake costly optimization of applications and system software to take full advantage of current ccNUMA architectures. Sun’s aim is to enable applications to take advantage of memory located on distributed systems by using Serengheti to ensure those systems behave for all intents and purposes as if they have a single store of shared memory. Serengheti – which reports say should be available in 1999 – is expected to combine techniques Sun has developed internally, plus those from its Cray superserver business and from Thinking Machines and Kendall Square Research. They include COMA, a Cache-Only Memory Architecture rival to ccNUMA that uses multiple levels of large caches, rather than a single large memory. A single node of a COMA machine would use local memory as a cache of a greater shared virtual memory, which is controlled by a single instance of the operating system.
Cheetah
Serengheti, or what’s been dubbed third-generation, 3D NUMA, is seen as supporting nodes with up to 30 processors, by that time on Sun’s next-generation 500MHz UltraSparc III RISC, codenamed Cheetah. Although Sun is convinced that further development of its shared memory architecture deployed in future generations of its existing SMP servers will accommodate the majority of users’ processing requirements, Serengheti is expected to be flexible enough to enable OEMs and customers to configure multiple systems in a variety of distributed arrangements. Sun has even built a conventional ccNUMA system in its labs, but says it has no plans to productize this particular solution, having built it primarily to test concepts. Sun claims it could have brought ccNUMA systems on to market long before Sequent or Data General, insisting those companies had little choice but to develop ccNUMA system product lines to create a market and stay in business. If ccNUMA’s so good Sun wonders, why isn’t Sequent or Data General recording industry-leading TPC-C performance numbers? Meantime, Sun’s about to announce a new set of technologies for clustering groups of its existing servers together – where each server is effectively and independent processing engine having its own copy of the operating system – as well as some of the first pieces of its Full Moon clustering software are offered as extensions to its Solaris Unix (CI No 3,118). Sun’s already promised Java and browser-based monitoring for four-node clusters, support for multiple logical hosts and high-availability versions of its WebServer and internet mail server for this year.