On Thu, 20 Jan 2000, Peter Rival wrote:
> No offense, but it is this type of thinking that will keep Linux
> out of the datacenter. What you must say it is not a typical (or
> realistic) workload _for me_. Hundreds of tasks is trivial here -
> we have systems running with well over 100 users actively working
> that are two or more generations old (that's a much bigger thing
> for Alpha than for Intel). On our newer systems we fully expect
> hundreds, if not thousands, of tasks.
I've seen machines like this in action. In one particular
case it was a university machine with 200 to 300 users
and usually somewhere between 1000 and 3000 processes.
The number of processes _on the runqueue_ at each point
in time, however, usually never got past 10...
What were the other processes doing?
- waiting for user input
- waiting for network or disk data
- waiting for a timeout
- just idling around for a longer time
- ...
If you have 100 people reading their email and each
email client needs 0.1% (that's a lot) of the CPU,
then you'll have, on average, just 0.1 (!) processes
on the runqueue...
You need to throw in a few compiles, scientific calculations
and people doing other cpu intensive stuff to get the load
average up to 8, on a machine with 300 users!
If there is anything to optimise for, it is this situation.
At any point in time (under these workloads) most processes
in the system will be idling; we'll want to do something to
make interactive response under load better.
Reduction in scheduling overhead for this case could be done
by moving some of the priority recalculation code, but that
will only work when most of the processes stay inactive for
more than several seconds. We can only determine what type
of optimisation we want to do once we have information on
the sleep/wake patterns of processes...
regards,
Rik
-- The Internet is not a network of computers. It is a network of people. That is its real strength.- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sun Jan 23 2000 - 21:00:26 EST