In case you are interested, I found why CURRENT performed so badly.

It turns out that CURRENT still does not have the fix to M_LEADINGSPACE
that permits writing into non-shared mbufs.  This caused the header
of forwarded packets to be pulled up in a separate buffer, and
triggered a known (to me at least!) performance bug in the 21143
which wastes about 5us when a packet is split across two descriptor.

With the one-line change shown below, forwarding speed goes
up to the same values as STABLE.

The change below has been committed to STABLE 7 weeks ago, but did
not go into CURRENT because there was some disagreement on the
semantics of M_LEADINGSPACE. However I would strongly vote for
committing this change to CURRENT as well, given the huge performance
implications (even if the 21143 were not buggy, not being able to
write into clusters hurts a lot of pieces of the networking stack).


Index: mbuf.h
RCS file: /home/xorpc/u2/freebsd/src/sys/sys/mbuf.h,v
retrieving revision 1.85
diff -u -r1.85 mbuf.h
--- mbuf.h      2001/09/30 01:58:35     1.85
+++ mbuf.h      2001/12/14 09:46:28
@@ -360,7 +360,7 @@
 #define        M_LEADINGSPACE(m)                                              \
        ((m)->m_flags & M_EXT ?                                         \
-           /* (m)->m_data - (m)->m_ext.ext_buf */ 0 :                  \
+           (M_WRITABLE(m) ? (m)->m_data - (m)->m_ext.ext_buf : 0):     \
            (m)->m_flags & M_PKTHDR ? (m)->m_data - (m)->m_pktdat :     \
            (m)->m_data - (m)->m_dat)

On Wed, Dec 12, 2001 at 10:42:06PM -0800, Luigi Rizzo wrote:
> Hi,
> I am testing the forwarding performance of CURRENT vs. STABLE
> (both more or less up to date, unmodified, with the latest performance
> patches to the "dc" driver, which I am using) and I am having some
> surprises.
> STABLE can forward approx 125Kpps, whereas CURRENT tops at approx 80Kpps.
> This is on the same hardware, 750MHz Athlon, fastforwarding enabled,
> a 4-port 21143 card, one input driven with a stream of up to 148Kpps
> (64 bytes each).
> Ability to transmit seems roughly the same (in both cases 138Kpps),
> and lack of CPU does not seem to be the problem (at least for
> CURRENT), so I am suspecting some difference in the initialization
> of PCI parameters, such as burst size etc, but I am unclear on
> where to look at.  Any ideas ?
>       cheers
>       luigi
> ----------------------------------+-----------------------------------------
>  Luigi RIZZO, [EMAIL PROTECTED]  . ACIRI/ICSI (on leave from Univ. di Pisa)
>  . 1947 Center St, Berkeley CA 94704
>  Phone: (510) 666 2927
> ----------------------------------+-----------------------------------------

To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-current" in the body of the message

Reply via email to