AltQ throughput issues (long message)

[Luigi Rizzo]

On Sat, Jul 31, 2010 at 09:10:04AM -0700, Patrick Mahan wrote:
...
> >part of it can be explained because AltQ counts the whole packet
> >(eg. 1514 bytes for a full frame) whereas iperf only considers the
> >UDP payload (e.g. 1470 bytes in your case).
> >
> 
> Okay, but that only accounts for 3% and I am seeing around 11%, any
> idea what might be accounting for the remaining 8%?

no, sometimes the extra icmp traffic plays a role, sometimes
it is just the shaper that is not precise and has systematic
errors (due to rounding in computing intervals and delays).
I cannot comment precisely on AltQ because i don't know
enough about its internals.

> >The other thing you should check is whether there is any extra
> >traffic going through the interface that competes for the bottleneck
> >bandwidth. You have such huge drop rates in your tests that i
> >would not be surprised if you had ICMP packets going around
> >trying to slow down the sender.
...
> Where do you see the drop?  If you are looking at the end of the pfctl 

iperf reports show that it is dropping tons of packets (see below, Lost/Total)

    > >>npx8# iperf -c 172.16.13.10 -p 7788 -u -b 25M -u -i 30 -t 200
    > >>------------------------------------------------------------
    > >>Client connecting to 172.16.13.60, UDP port 7788
    > >>Sending 1470 byte datagrams
    > >>UDP buffer size: 9.00 KByte (default)
    > >>------------------------------------------------------------
    > >>[  3] local 172.16.13.60 port 7788 connected with 172.16.38.80 port 41064
    > >>[ ID] Interval       Transfer     Bandwidth       Jitter   Lost/Total  
    > >>Datagrams
    > >>[  3]  0.0-30.0 sec  5.96 MBytes  1.67 Mbits/sec  0.710 ms 59453/63706 
    > >>(93%)
    > >>[ ID] Interval       Transfer     Bandwidth       Jitter   Lost/Total
    > >>Datagrams
    > >>[  3] 30.0-60.0 sec  5.95 MBytes  1.66 Mbits/sec  0.736 ms 59616/63859
    > >>(93%)


> >
> >BTW have you tried dummynet in your config?
> >
> 
> How would you suggest using dummynet?  Is it workable for a QoS solution?

Very workable. see http://info.iet.unipi.it/~luigi/dummynet/
especially the video/slides at the top of the page.

cheers
luigi