Despite having its own serial link implementation S3.MP (CI No 2,738) and the patents from failed massively parallel house Kendall Square Research Inc in hand, it seems that Sun Microsystems Inc is shying away from swallowing a whole cc-NUMA cache-coherent Non-Uniform Memory Architecture a la Data General Corp and Sequent Computer Systems Inc in favor of what it is calling an ‘incremental adoption’. The critical element for a NUMA-type implementation, according to Jeff Rulifson, director of technology development at the company’s Sun Microsystems Computer Co hardware arm, are the RAS reliability, availability, servicability characteristics of the operating system and the tuning of the architecture for efficiency. There’s a difference between making it [NUMA] work and demonstrating it can’t fail, he said. Instead Rulifson expects the technology to trickle into future versions of Solaris over time, although the strategy isn’t set in stone. We can’t yank out the kernel and guarantee back-wards-compatibility, he admitted, and having been through the experience once before, we have no desire for a repeat, referring to the Motorola Inc 68000 to Sparc transplant in 1988. Sun is also concerned about the scalability of its serial link and the way the interconnect controller, S-Connect, would perform. Sun has added what it tags a simple COMA implementation to its original cc-NUMA architecture including typical COMA-type dynamic data migration and replication, but with reduced hardware and protocol complexity. The key to the technology, according to Sun, is the use of the memory management unit on a commodity processor to build the COMA cache or ‘attraction memory’. A compute node of a simple COMA machine uses local memory as a cache of a greater shared virtual memory – a cache in which space is allocated and reclaimed at a page granularity, but which is fully associative. Each memory page is sub-divided and a state is attributed to it. A hardware protocol controller controls processor accesses to these sub-divisions with responsibility for maintaining fine grain coherence. The operating system, through interaction with the memory manager, becomes responsible for co-ordinating allocation and replacement of data space in the attraction memory. The migration of complexity from hardware to software, claims Sun, simplifies the memory system and protocol requirements and enables more flexible cache management strategies. Quite where the Kendall Square stuff plays, if at all, remains unclear. Sun has a bunch of other potential clustering mechanisms in hand, including Dolphin Interconnect A/S’s Scalable Coherent Interface-based interconnect for clustering its commercial servers (CI No 2,836), Fiber Channel and the Asynchronous Transfer Mode Global-Works parallel software Thinking Machines Corp has fitted to its UltraSparc servers.
