DP performance

Danial Thom danial_thom at yahoo.com
Wed Apr 27 08:46:34 PDT 2005


--- Matthew Dillon <dillon at xxxxxxxxxxxxxxxxxxxx>
wrote:
> 
> :
> :A machine running 1.2 with one or two opterons
> :goes into livelock at just about the same
> traffic
> :level, which is about the same behavior as
> with
> :older freeBSD versions. Are such things still
> a
> :work in progress?
> :
> :Danial
> 
>     The network path is still operating under
> the big giant lock.
>     We expect to get it out from under the BGL
> this year.
> 
>     Could you explain the characteristics of
> the livelock you
>     are observing ?  Interrupt layer livelock
> shouldn't happen
>     with a modern network adapter, but it is
> certainly possible
>     that other forms of livelock are effecting
> the system.  The
>     biggest single problem is still the BGL but
> I am also rather
>     unhappy with our current scheduler.

I wonder why/how you say that? Interrupt livelock
has always been a way of life with FreeBSD 4.Its
fairly easy to reach while bridging with a large
number of firewall rules with an fxp or em
driver.
Just increase the rate until you lose the 
keyboard. What is your idea of a "modern"
adapter?

Is there a way to direct specific interrupt 
activity to a single processor, or to tune the
behaviours in any way? It seems absurd that there
is no networking benefit to adding another
processor two years into the project.

Danial

__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 
http://mail.yahoo.com 





More information about the Users mailing list