Data General Corp is pretty pleased that Sun Microsystems Inc bothered to rubbish its ccNUMA architecture when heralding Sun’s own Serengheti shared memory technologies (CI No 3,253), because it reckons two years ago, Data General was not even on Sun’s radar screen. Sun made a number of comments regarding ccNUMA, including the fact that it doesn’t want its new architecture sullied by association with the cache-coherent distributed shared memory systems such as that offered by Data General and Sequent Computer Systems Inc, which Data General is keen to refute. Sun claims ccNUMA is fundamentally flawed because of a latency ‘bug’, or time penalty for accessing remote memory. Johan Edward, a technology marketing manager for Data General, says ccNUMA in fact has extremely low latency. Latency, or a delay, occurs when a symmetric multiprocessing node needs to access remote memory. Edward says in practice, this happens very rarely anyway, because everything in the system is optimized to insure the processor has minimal requirement to go out to the remote memory. For a start, the operating system, in this case DG-UX, has ‘affinity’, which works out which processes are likely to access which memory and attempts to keep everything on one symmetric multiprocessing board. In addition, Data General uses coherent second level cache memory, which insures that most of what is needed is available in local cache. When remote memory does need to be accessed, it is done via a dedicated memory bus, and Edward claims there is only a small time penalty for doing this, and it occurs very rarely anyway. Data General claims ccNUMA works in exactly the way Sun describes its COMA Cache-only Memory Architecture, with multiple levels of large caches rather than a single large memory. As for Sun’s claims that it could have developed a ccNUMA system long before Data General or Sequent, Data General points out that it has been actually shipping ccNUMA systems since August 1995, and that if Sun could have got there first, why didn’t it, and why won’t Serengheti appear before 1999? Sun also asked why, if ccNUMA was so good, Data General hadn’t published any TPC-C performance figures. Data General says it has consistently won in customer benchmarks with its Aviion NUMA servers, and it has surpassed its own expectations for scalability, recording 95% linear scalability in what it believes is the world’s largest ccNUMA system, recently installed for UK supermarket chain J Sainsbury Plc, with a 32 processor, (eight nodes of four processors) system. It also says it beat MPP Massively Parallel Systems in a customer pitch, where the MPP system cost $30m compared to Data General’s $3m offering. However, what the company is really waiting for, it says, is to publish performance figures for its recently announced fiber channel interconnect. Edwards claims the performance of the combination of ccNUMA and fiber channel is blistering. He is talking about input/output throughput of 560 megabytes per second, compared to the 100Mbps achieved even with the Sainsbury system, and performance of around 30,000 transactions per second, for somewhere around a tenth of the cost of massively parallel systems. Data General says some applications can be tuned and optimized for ccNUMA, and indeed Oracle Corp’s Oracle 8 has NUMA tuning options available within it, but it maintains that this is just a bonus. Applications will run untouched with ccNUMA and excellent scalability is achieved without any tuning at all, Edwards says. Ultimately, the company says its systems are aimed specifically at the high end commercial market, and it believes it delivers excellent price performance for this market. Sun’s suggestion that Data General and Sequent had little choice but to develop ccNUMA product lines to create a market and stay in business (CI No 3,253) makes Edwards laugh with delight. We think in business its a jolly good idea to create a market and stay in business.
