By William Fellows
Students at six US universities including MIT are test driving version 1.0 of Sun Microsystems Inc’s distributed shared memory technology running on Ultra Enterprise 6000 servers connected with network interface cards Sun that has developed to implement its ccNUMA. ComputerWire has learned that Sun, which already has a second cut of the technologies running in house, will probably begin introducing the technology in 1999. If it can, it will do it in such a way that users and ISVs will not know any different. The key, it says is to extend the current SMP symmetric multiprocessing model in which all processors are created equally to memory, such that all memory is created equally. It says it categorically will not break the SMP model like ccNUMA server vendors Silicon Graphics Inc or Sequent Computer Systems Inc have done by introducing latency ‘bug’. Notice it says, how quiet Hewlett-Packard Co has been about the ‘bug’ it’s introducing to extend the V-Class to 128-ways. Sun says its ccNUMA version 1.0 latencies are of cache misses below Sequent’s one-in-five or one-in-ten but have been deliberately kept high enough for students to be able to test and examine behavior clearly. With version 2.0 of the technologies, Sun says it believes that latencies are now at the point at which they don’t matter to applications, in much the same way that programs running on current SMP machines do not care about the latencies associated with accessing L1 and L2 caches. As such, Sun says the ccNUMA technology it eventually introduces will not require users or ISVs to alter their applications in any way. That’s why it has so much trouble articulating its distributed shared memory strategy. As far as it’s concerned, it’s not going to break the SMP model like other vendors. It says remarks made at a recent analyst meeting in Rome about categorically not doing ccNUMA (CI No 3,593) set off all kinds of alarm bells internally and concretized the extent of the problem. The aim of ccNUMA is extensibility first and foremost. Sun does not expect the work to be used to create more massive systems in the first instance but will allow users to add processing power incrementally. Nevertheless, he who has the fattest node wins, Sun believes. The SMP model is being built out to its limits. ccNUMA takes over from there. Sun says it clearly has the fattest node. It’s in no rush to introduce ccNUMA because the E100000 Starfires give it a lot of room to maneuver. A lot more than HP, Sequent or SGI, which have all already been burnt by ccNUMA. Sun says the purchase of the Cray SuperServer business from SGI actually gave it the potential to get ccNUMA technology to market very quickly if it had so chosen. It admires SGI’s ccNUMA-based Origin server work, but believes the company is getting killed by the way it swapped out conventional SMP for a ccNUMA model that is still not complete. Sun says that it has been working on several forms of ccNUMA technologies for two or three years. The engineering team it hired from Thinking Machines Corp and technology from defunct Kendall square Research plus other Sun divisions have all contributed work. Indeed some preparatory work has already been delivered in the new Solaris 7 release it says, with more to follow in Solaris 8.