SMP and dual tulip: Solution!
Tue Nov 9 21:49:00 1999
On Tue, Nov 09, 1999 at 03:41:04PM -0800, Brian Macy allegedly wrote:
> Note that this is likely a bad "solution". Your memory bandwidth is
> likely halved... don't know why, it just happens.
> Brian Macy
Erm, hmm, it shouldn't. I must admit I haven't done any test with
regard to memory bandwidth, but I'll check. The memory bandwitdh
you measure should be the same with both one and two cpu's (pretty
obvious). It probably reaches it's peak at 300-400 MB/sec (if it
even gets that high). The MPS version used shouldn't have any
effect on the memorybandwith. Memorybandwith on Intel based systems
is crappy anyway. I guess a crossbar-switch like the ones used in
the more high-end Sun, DEC and SGI boxes is probably too expensive.
I do wonder what kind of solution they've come up with for four and
eight CPU Xeon boards - 400 MB/sec is just no enough for eight CPU's.
They've probably dug up multiplexing of memorybanks again :)