>The 512CPU SGI thing I was talking about is a Shared Memory NUMA >system running single system imag e, which is different from MPP or a >clustered system. It runs upto 512 CPUs in a single node.
I think most of us are well aware of which system is meant. >If a kernel scales well with single image on such a system, it can >very well scale (upto the point where it comes to hardware limitations) >on 512 CPU SMP box if it was available. For some meaning of "scales well". These systems are typically used as "HPC" systems; they don't require much from the kernel by way of scaling as such jobs are mostly userland, CPU bound. Doing that on a N-way box is trivial, for nearly any N. "Scaling" is not a property with a single parameter; it's a property of a workload. E.g., the TPC-C benchmark scales on clusers because there isn't much inter node traffic. But other jobs don't run at all well (worse on two or more nodes than on one) because the requirements of the job are not met. Similarly, having Linux running on a 512-CPU box doesn't mean it "scales" to 512 CPUs for any work load, just for some. >And NASA runs such a box in production, with Linux, for "Mission Critical" >stuff. Ah, yes, NASA. Casper _______________________________________________ opensolaris-discuss mailing list [email protected]
