Discussions have been continuing on and off for at least a year between Microsoft Corp and supporters of Non Uniform Memory Architecture-based systems, and despite Data General Corp saying it’s a lost cause – at least for this year – Sequent Computer Systems Inc is soldiering on, determined to try to convince Microsoft to adopt NUMA, an architecture Microsoft has regularly dismissed, in order to get Windows NT to scale above eight processors. As ClieNT Server News reports, not only must Sequent persuade Microsoft into the NUMA camp, it wants it to adopt its own particular brand of the architecture. To do this, the company is set to build a 12-way shared-everything NUMA-Q prototype, the minimal size for a NUMA configuration, it says, and run NT across it. Sequent has already spent $54m on NUMA-Q for its Unix systems over the last four years, and expects to spend another $20m this year, so it wants to get value for its money. But to get NT to run efficiently on NUMA it looks as if it will require significant changes in the operating system, and according to Sequent, nobody knows exactly what those changes would be. It plans to see what it can do from the hardware angle with the Hardware Abstraction Layer, and then turn the system over to Microsoft and see if it could be persuaded to change the NT kernel.
Academic exercise
Data General is considering a similar kind of academic exercise as Sequent, but not until it gets its Unix-on-NUMA boxes out of the door some time this spring. It has already done some limited testing on NT, since it can’t be ignored, and it would be interested to see what base line performance Sequent gets with the current version of NT on its prototype and how much better that is, if at all, than a four-way system. Data General believes it won’t need to build a special NUMA prototype for NT since NT already runs on its nodes, which are built out of standard Intel Corp motherboards. It says a better approach might be to talk to Microsoft not about NUMA per se but about the components that would make NT NUMA-compatible, such as process affinity and the size of memory NT supports. In the long term, Data General is counting on market realities, that is the success with NUMA on Unix-based systems, to persuade Microsoft of its worthiness. Meanwhile, Sequent claims NUMA is necessary for NT as a data center system, a way to overcome the limitations of big bus systems, which are particularly obvious with the amount of data the Intel P6 processor can deliver. Sequent believes it would be most sensible for Microsoft to be able to support NUMA by the time SQL Server 7 arrives in the middle of 1998. Microsoft, however, may be planning to address the problem by using NCR Corp’s OctaScale technology, a NUMA hybrid that simplifies the complex way NUMA handles memory, and runs NT well up to its own eight-processor limit. Inside Microsoft, there is a quiet enterprise effort called the Scalability Project, said to be geared to OctaScale and SQL Server 7. Early benchmark simulations suggest OctaScale systems will be able to achieve better than 11,000 transactions-per-minute, whereas the best of the current NT four-way systems achieve 7,500 transactions-per-minute.
