You're actually comparing scalability. However, I would not make the comparison on the terms of these machines all being SMP. You actually have a mixture of different models for achieving application-specific scalability:
SMP - Sun
NUMA - SGI Origins
MPP - T3E
SMP, by definition, has all processors sharing a common memory (uniform memory), with memory contention/bandwidth typically becoming the limiting factor on scalability.
NUMA utilizes cache coherency strategies to add multiple memory pools, introducing the concepts of local (memory local to one or more processors, e.g. those on the same board) and far (non-local memory owned by other processors in the machine) memory accesses. NUMA reduces the contention for memory, thus achieves a higher scalability with the performance cost being in terms of memory latency on far accesses.
Lastly, MPP is essentially NUMA without cache coherency, which is why programming MPP machines requires additional effort on the part of application developers.
Certain types of applications favor different scalability models, so one machine does NOT fit all. For general purpose computing, SMP is typically the best bang for the buck, since the architectural complexity is typically less. However, with the right problem to solve, NUMA and MPP systems will fly past SMP, but typically cost way more $$$ to get there.
Solaris probably cannot improve scalability much further with SMP with today's hardware. They are utilizing crossbar memory backplanes in the highend already to get to 64 processors. Bottom-line, I believe they are doing a darned good job on the SMP hardware they produce.
I submitted the same story a couple hours before your time and it was also rejected. I wonder why? Perhaps they had to get NSA approval to post the story.....
Slashdot Mirror
SMP, by definition, has all processors sharing a common memory (uniform memory), with memory contention/bandwidth typically becoming the limiting factor on scalability.
NUMA utilizes cache coherency strategies to add multiple memory pools, introducing the concepts of local (memory local to one or more processors, e.g. those on the same board) and far (non-local memory owned by other processors in the machine) memory accesses. NUMA reduces the contention for memory, thus achieves a higher scalability with the performance cost being in terms of memory latency on far accesses.
Lastly, MPP is essentially NUMA without cache coherency, which is why programming MPP machines requires additional effort on the part of application developers.
Certain types of applications favor different scalability models, so one machine does NOT fit all. For general purpose computing, SMP is typically the best bang for the buck, since the architectural complexity is typically less. However, with the right problem to solve, NUMA and MPP systems will fly past SMP, but typically cost way more $$$ to get there.
Solaris probably cannot improve scalability much further with SMP with today's hardware. They are utilizing crossbar memory backplanes in the highend already to get to 64 processors. Bottom-line, I believe they are doing a darned good job on the SMP hardware they produce.
I submitted the same story a couple hours before your time and it was also rejected. I wonder why? Perhaps they had to get NSA approval to post the story.....