Re: [Announce] [patch] Modular Scheduler Core and Completely Fair Scheduler [CFS]

[Willy Tarreau]

On Sat, Apr 14, 2007 at 12:30:17AM +0200, Ingo Molnar wrote:
> 
> * Daniel Walker <dwalker@xxxxxxxxxx> wrote:
> 
> > I'm not in love with the current or other schedulers, so I'm 
> > indifferent to this change. However, I was reviewing your release 
> > notes and the patch and found myself wonder what the logarithmic 
> > complexity of this new scheduler is .. I assumed it would also be 
> > constant time , but the __enqueue_task_fair doesn't appear to be 
> > constant time (rbtree insert complexity).. [...]
> 
> i've been worried about that myself and i've done extensive measurements 
> before choosing this implementation. The rbtree turned out to be a quite 
> compact data structure: we get it quite cheaply as part of the task 
> structure cachemisses - which have to be touched anyway. For 1000 tasks 
> it's a loop of ~10 - that's still very fast and bound in practice.

I'm not worried at all by O(log(n)) algorithms, and generally prefer smart log(n)
than dumb O(1).

In a userland TCP stack I started to write 2 years ago, I used a comparable
scheduler and could reach a sustained rate of 145000 connections/s at 4
millions of concurrent connections. And yes, each time a packet was sent or
received, a task was queued/dequeued (so about 450k/s with 4 million tasks,
on an athlon 1.5 GHz). So that seems much higher than what we currently need.

Regards,
Willy

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/