Petz and all, re: AMD's point-to-point bus and multiprocessor performance,
After doing a little research into the subject, I learned a little something interesting about the K7's point-to-point bus. In multiprocessor systems, compared to a shared processor bus like Xeon, the K7 actually has longer latency, even though data bandwidth is greater. The reason has to do with cache coherence. I'll avoid the gory details, but suffice to say that with a shared multiprocessor bus, every processor can watch the transactions go by on the bus and snoop accordingly. With point-to-point connections, if a processor issues a memory access, all the other processors must be notified separately before the memory access can go through.
That's one advantage of a shared multiprocessor bus. So for two-to-four-way systems, Xeon will exhibit lower latencies than K7. On the other hand, for eight way systems and higher, the latency advantage will probably disappear because the Xeon will have to resort to multiple buses which need to "snoop" each other, thereby adding latency.
It'll be interesting to see how well multiprocessor K7 servers run on server benchmarks like TPC, especially with the point-to-point connections that isn't the panacea that AMD makes it out to be.
Tenchusatsu |
