On Thu, Apr 28, 2016 at 3:17 PM, Rich Pieri <[email protected]> wrote: > On 4/28/2016 4:28 AM, Bill Bogstad wrote: >> So memory was shared? between the QBBs? This sounds more like a NUMA >> architecture environment. What would you say are the differences >> between this definition >> of SSI and NUMA? > > In a NUMA machine, memory is directly attached to the CPUs but not all > of that memory is local to each CPU. Galaxy wasn't NUMA. Each QBB was > NUMA: 4 processors with 512MB local to each with the rest being > non-local but still directly attached. Memory in one QBB was not > directly attached to the processors in other QBBs; it was shared via > software over a fibre data bus.
Can you clarify what you mean by "shared via software"? Did the virtual memory system, page fault data from remote QBBs as needed or was there a fibre bus transaction every time a local process accessed remote memory? I understood your original note to mean that from a programs perspective it could allocate/use memory in other QBBs transparently (except for possible performance differences). The fact that this was done via a fibre data bus vs. a faster local bus would seem to me to be an implementation detail. It still sounds like a NUMA with three levels of memory access (CPU local, QBB local, remote QBB). But all of the memory transparently visible in a program's address space. Bill Bogstad _______________________________________________ Discuss mailing list [email protected] http://lists.blu.org/mailman/listinfo/discuss
