lkml.org 
[lkml]   [2002]   [Mar]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: Futexes III : performance numbers
Date
On Wednesday 06 March 2002 02:54 am, Rusty Russell wrote:
> On Tue, 5 Mar 2002 16:23:14 -0500
>
> Hubertus Franke <frankeh@watson.ibm.com> wrote:
> > Interesting... the strict FIFO ordering of my fast semaphores limits
> > performance as seen by 99.71% contention, so we always ditch
> > into the kernel. Convoy Avoidance locks 2.5 times better.
> > Wohh futex rock, BUT... with 0.29% contention it basically tells
> > me that we are exhausting our entire quantum getting the lock
> > without contention. So their is some serious fairness issue here
> > at least for the tightly scheduled locks. Compare the M numbers
> > for 2 and 3 children.
>
> Fairness <sigh>. This patch should be much more FIFO: it works by handing
> the mutex straight to the first one on the queue if there is one, and only
> actually "freeing" it if there's noone waiting.
>
> Unfortunately, it seems to hurt performance by 50% on tdbtorture (although
> there are weird scheduler things happening too).
>
> Here's the "fair" patch:
> Rusty.

Rusty why not provide something along the line <snipped all over the place>.

#define FUTEX_DOWN (0)
#define FUTEX_UP (1)
#define FUTEX_FAIR_UP (2)

static int futex_up(struct list_head *head, atomic_t *count)
{
atomic_set(count, 1);
smp_wmb();
wake_one_waiter(head, count);
return 0;
}

static int futex_fair_up(truct list_head *head, atomic_t *count)
{
spin_lock(&futex_lock);
if (!pass_futex(head, count))
/* Noone to receive: set to one and leave it free. */
atomic_set(count, 1);
spin_unlock(&futex_lock);
return 0;
}


asmlinkage int sys_futex(void *uaddr, int op)
{
<..... snip ....>

head = hash_futex(page, pos_in_page);
switch (op) {

case FUTEX_DOWN:
ret = futex_down(head, page_address(page) + pos_in_page);
break;

case FUTEX_UP:
ret = futex_up(head, page_address(page) + pos_in_page);
break;

case FUTEX_FAIR_UP:
ret = futex_fair_up(head, page_address(page) + pos_in_page);
break;

default :

<..... snip ....>
}


This would satisfy the fair vs. calock issue, you let the app decide what to
use. Best of all, it seems to me you can even mix it.
Imagine, if a process knows it will soon reaquire the lock, then it would
use FUTEX_UP to avoid being tagged back to the end of wait queue
avoiding costly scheduling events. At the same time, if the process knows
that its done for a while with that lock, then it issues FUTEX_FAIR_UP.
The best of two worlds..

It also shows how cleanly the code can be expanded in the future.
The more I look at the hash queues the more I like it.
Will look at the rwlock now. Let me know what you think.

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

\
 
 \ /
  Last update: 2005-03-22 13:24    [W:0.200 / U:0.416 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site