[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Linux-QoS overhead (was Re: RH6.0&6.1 and diff-serv)

>>>>> "Werner" == Werner Almesberger <almesber@lrc.di.epfl.ch> writes:
    >> Another question. What load the PPro-200 can handle.  10 cbq-classes x
    >> 19200bps ?! what is diff-serv overhead ?!

    Werner> A few dozen Mbps should be okay, unless you have excessively
    Werner> complex classifiers or such.

  I've measured an Amd K6-200 do approximately 30Mb/s in software using three
input streams of 100Mb/s and one outgoing 100Mb/s NIC.

  I had configured 90Mb/s of outgoing bandwidth, and offered it up to
300Mb/s of incoming data. I will attach the setup script below. I can
provide the SmartFlow 1.0 config file if someone has the required test

  Using hardware assist for the classifiers, I was able to get that up to
around 65Mb/s. (It was 80Mb/s on an K6-333). This was with 2.2.10 plus
the patches at the time (last August). We showed this demo to some select
people at Interop 1999 in Atlanta. I don't think we will bring it to Las
Vegas, since we won't have space for it in the booth. Maybe.

  Having offloaded the classification problem into O(1) hardware, the
bottleneck appears to be in the outgoing scheduling. I think the major 
challenge for the Linux-QOS stuff is making sure that it will integrate
seemlessly with hardware assist devices. Incoming is pretty easy, but
outgoing is where the problem is. I know that 2.3/2.4 revises a bunch of
this. I wish we had time to cook up some better outgoing assist. Perhaps
I'll have to learn VHDL afterall.

  I hope to try this again with 2.4. I haven't had time to try 2.3 at all.
I expect real improvements from what I've read.

  I also expect to repeat this test with NetBSD 1.5 in the spring. (1.5
will have ALTQ+KAME in it. I know that ALTQ has a different driver model)