> http://www.theregister.co.uk/2013/04/17/state_of_linux_2013/ > > Some interesting points from an HPC point of view: > > > "For example, for some kinds of workloads, the NUMA (non-uniform memory > access) problem is all-important. This is particularly true of distributed > application clusters. A long as a CPU in a cluster node is working on data > that exists locally, within the node's own memory or storage, processing > is fast. But as soon as that CPU needs to access data that exists on some > other node in the cluster - data that must pulled across some slow type of > transport, like an Ethernet link - performance can degrade rapidly. Some > Linux kernel developers would like to improve the CPU scheduler to handle > such situations better."
I am puzzled by this paragraph. I don't follow the kernel development list anymore, but is there some kind of effort to support shared memory over Ethernet? I think "degrade rapidly" is a bit of an understatement. -- Doug > -- Doug -- Mailscanner: Clean _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
