On 01/19/2006 07:41:54 PM, Kevin Closson wrote:
You'll make me feel like an old man now. With Legacy iron, everyone accept Sequent, threw HW interrupts out on the single bus. Sequent SMPs (Symmetry) has a secondary bus for the sole purpose of distributing interrupts... so other systems had what we always called pregnant CPUs, because they were the unlucky ones that always snacked up the interrupts...
I don't know how old are you, but I seem to be an even older man. I remember something called ASMP (VAX 782, VMS 4.4, 4.6 and 4.7, the one which had "null process") in which only one processor was allowed to run in the kernel mode. That meant that all system services were handled by a single CPU.
and if you were some poor Oracle shadow process with affinity to that CPU, uh, well, let's just say that you get an unfair amount of starvation...and oh my the chaos that ensues when processes held latches on the "pregnant CPUs". The was why Sequent stomped all other SMPs. Other vendors' benchmarking folks would have to go in there and set hard affinity of Oracle processes to higher order CPUs so that this sort of pathology wouldn't occur.
It was DEC who was the first one with the real SMP. VAX 6200 with then revolutionary XMI bus was the first true SMP box and it came about in 1990 or so. VAX 9000 went step further and introduced memory mapped I/O ("I/O channels" in the mainframe terminology). Unfortunately, when that was created, DEC was already practically dead. DEC marketing was so atrocious that it managed to bury the company that had by far the best OS in the world, far superior to almost anything that exists today. HP also blew it majestically. There was an attempt to introduce something called I2O on Linux and Unix systems, but the protocol was closed and nobody was really interested. Shame.
-- Mladen Gogala http://www.mgogala.com -- //www.freelists.org/webpage/oracle-l