tcp_output starving -- is due to mbuf get delay?

Jin Guojun [DSD] j_guojun at lbl.gov
Wed Apr 9 17:23:49 PDT 2003


The interesting result --

In normal FreeBSD TCP stack, the large delay goes away, but more small delays.

Apr  9 17:10:01 tcp_lion /kernel: sosend: td 23 3424 (bumped up from 1920)

The performance dropped badly to 92 Mb/s (no loss)

In my new TCP, I did not see any delay above 5 us, which is good, but
the overall TCP performance drop to 120 Mb/s. So, maybe either
there are a lot delay below 5 us and about 100 ns, or some other bottleneck
is trigged somewhere.

I guess there is more work to do to determine what is going on :-(
I will post whatever I will discover.

Thanks for pointing to the patch.

    -Jin

Sean Chittenden wrote:

> > When testing  GigE path that has 67 ms RTT, the maximum TCP throughput is
> > limited at 250 Mb/s. By tracing the problem, I found that tcp_output() is
> > starving
> > where snd_wnd and snd_cwnd are fully open. The snd_cc is never filled beyond
> > the 4.05MB even though the snd_hiwat is 10MB and snd_sbmax is 8MB. That is,
> > sosend never stopped at sbwait. So only place can slow down is the mbuf
> > allocation
> > in sosend(). The attached trace file shows that each MGET and MCLGET takes
> > significant time -- around 8 us at slow start time, and gradually increasing
> > after that
> > in an range 18 to 648.
> > Each packet Tx on GigE takes 12 us. It average mbuf allocation takes 18 us, then
> >
> > the performance will be reduced to 40%, in fact it is down to 25%, which means
> > higher average delay.
> >
> > I have change NMBCLUSTER from 2446 to 6566 to 10240, and nothing is improved.
> >
> > Any one can tell what factors would cause MGET / MCLGET to wait?
> > Is there any way to make MGET/MCLGET not to wait?
>
> Luigi posted a patch about this a while back (last summer sometime,
> iirc).
>
> http://people.freebsd.org/~seanc/patches/#o1_mbuf_lookup
>
> I updated his patch but haven't had a chance to test it.  If you're
> feeling brave, see if applying this patch fixes this bottle neck.  -sc
>
> --
> Sean Chittenden

--
------------ Jin Guojun ----------- v --- j_guojun at lbl.gov ---
Distributed Systems Department          http://www.itg.lbl.gov/~jin
M/S 50B-2239                            Ph#:(510) 486-7531 Fax: 486-6363
Lawrence Berkeley National Laboratory,  Berkeley, CA 94720




More information about the freebsd-performance mailing list