On Wed, Aug 16, 2006 at 12:30:29PM -0400, Jeff Garzik wrote:
> Linas Vepstas wrote:
> >
> >The recent set of low-waterark patches for the spider result in a
>
> Let's not reinvented NAPI, shall we...
??
I was under the impression that NAPI was for the receive side only.
This round of patches were for the transmit queue.
Let me describe the technical problem; perhaps there's some other
solution for it?
The default socket size seems to be 128KB; (cat
/proc/sys/net/core/wmem_default) if a user application
writes more than 128 KB to a socket, the app is blocked by the
kernel till there's room in the socket for more. At gigabit speeds,
a network card can drain 128KB in about a millisecond, or about
four times a jiffy (assuming HZ=250). If the network card isn't
generaing interrupts, (and there are no other interrupts flying
around) then the tcp stack only wakes up once a jiffy, and so
the user app is scheduled only once a jiffy. Thus, the max
bandwidth that the app can see is (HZ * wmem_default) bytes per
second, or about 250 Mbits/sec for my system. Disappointing
for a gigabit adapter.
There's three ways out of this:
(1) tell the sysadmin to
"echo 1234567 > /proc/sys/net/core/wmem_default" which
violates all the rules.
(2) Poll more frequently than once-a-jiffy. Arnd Bergmann and I
got this working, using hrtimers. It worked pretty well,
but seemed like a hack to me.
(3) Generate transmit queue low-watermark interrupts,
which is an admitedly olde-fashioned but common
engineering practice. This round of patches implement
this.
--linas
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
