[geeks] Re: [rescue] Re: kernel scalability....

[geeks at sunhelp.org]

n2vip at yahoo.com writes:

>> However - why does running multiple unix instances under VM help?
>>
>> Is this a context-switch-to-hide-latency type effect, like barel
>> processors use?

>Dave,

>I suspect that the CPU is *so fast* that if the mainframe were to run a
>single thread, the relatively slow I/O would really bog down the process (as
>measured in *clock time*, not CPU time).

The ratio of CPU speed to IO is the exact opposite of that in traditional
mainframes. Compared to your typical workstation, they have wussy CPUs
and mind-numbing IO.

>In my (casual) experience, mainframes have an incredible ability to absorb
>more tasks without effecting clock time for a given task too much, much like
>a bus - it doesn't take longer to move 50 people from here to there than if
>there were only one person in the bus...

What you're seeing there is a "well balanced machine" as opposed to the
jetpack-strapped-on-a-tricycle that most people are used to. 

How far off Amdahl's Law are PCs these days?

>As i understand it, the *nix kernel does one thing at a time, and task
>switches based on time slices - in such an implementation of a "classic"
>*nix kernel on a mainframe, you would spend most of your time waiting for
>I/O devices. But this is a *real* simplification, but I think it holds true
>as details are added/model refined to closer map the actual tasks
>involved...

Perhaps I'm misunderstanding you here, but...

Unix does not hang waiting on IO. A process either gives up its
timeslice by making a request to the kernel, or it hits its time limit and
gets interrupted. A process does not get scheduled for CPU time if it's
waiting for IO.

-------- David Fischer --------- dave at cca.org --------- www.cca.org --------
--------- "I prefer the ridiculous to the sublime." - James Chance ---------