Re: Analysis for Linux-2.5 fix/improve get_pid(), comparing various approaches

[Hubertus Franke <frankeh@watson.ibm.com>]

On Friday 09 August 2002 11:36 am, Andries Brouwer wrote:

!!!!!!!!!!! You are in a different space !!!!!!!!
All work was done under the assumption of 16-bit pid_t.
I stated yesterday already that for NumTasks substantially smaller
than the pid_t supported size, this won't be a problem
as your analysis states and my data also states.

You have two choices 
(a) move Linux up to 32-bit pid_t
(b) stick within the current 16-bit discussion.

> On Fri, Aug 09, 2002 at 07:22:08AM -0400, Hubertus Franke wrote:
> > Particulary for large number of tasks, this can lead to frequent exercise
> > of the repeat resulting in a O(N^2) algorithm. We call this : <algo-0>.
>
> Your math is flawed. The O(N^2) happens only when the name space for pid's
> has the same order of magnitude as the number N of processes.
> Now consider N=100000 with 31-bit name space. In a series of
> 2.10^9 forks you have to do the loop fewer than N times and
> N^2 / 2.10^9 = 5. You see that on average for each fork there
> are 5 comparisons.
> For N=1000000 you rearrange the task list as I described yesterday
> so that each loop takes time sqrt(N), and altogether N.sqrt(N)
> comparisons are needed in a series of 2.10^9 forks.
> That is 0.5 comparisons per fork.
>
> You see that thanks to the large pid space things get really
> efficient. Ugly constructions are only needed when a large fraction
> of all possible pids is actually in use, or when you need hard
> real time guarantees.
>
> Andries


-- 
-- Hubertus Franke  (frankeh@watson.ibm.com)