[altq 1380] Re: Need help: PRIQ Details

["Tippyarat Tansupasiri" <fon@cs.ait.ac.th>]

Dear Nicolas,

What should be the size of your very very small packet?

I used "Iperf" program to gernate those UDP flows with 1470 byte datagram
size and 9 KByte buffer size (their default values). As stated in Iperf
document, Iperf "create a constant bit rate UDP stream" which is "similar to
voice communication".

All links used in the experiment are Ethernet (100Mbps and 10Mbps). The
result is based on 1 second statistic (command -> altqstat -w 1). I
understand the point that the monitoring window should be small, but 1
second is the least I can specified, right?

Should I try any other settings or traffic generators? Could you please
suggest?

Regards,
Tippyarat T.

----- Original Message -----
From: "Nicolas Christin" <nicolas@cs.virginia.edu>
To: "ALTQ Mailing-List" <altq@csl.sony.co.jp>
Sent: Friday, March 01, 2002 12:00 AM
Subject: [altq 1377] Re: Need help: PRIQ Details


> On Thu, 28 Feb 2002, Tippyarat Tansupasiri wrote:
>
> > How about UDP? Would the result be the same as TCP? All traffic stated
in
> > the previous mail are based on UDP.
>
> As I said, it depends on your traffic generator. For pure
> constant-bit-rate UDP with a very very small packet size (emulating
> fluid-flow CBR), what you should see is: high_pri gets all the bandwith,
> other classes get nothing.
>
> I suspect that you:
>
> a) either do not generate CBR traffic,
> b) generate large size packets which in fact create short bursts.
>
> PRIQ forwards from high_pri only if high_pri is backlogged. Turns
> out that this is not constantly the case in your example. As a rule of
> thumb, what you will generally see is that hgih_pri gets more throughput
> than low_pri, but PRIQ generally does not allow for quantification of the
> difference. In other words, what you see is normal.
>
> One last thing: the throughput is always a time average, there is
> no such thing as "instantaneous throughput". That's an oxymoron, it would
> require a division by zero. So you want to pay close attention to the
> monitoring window you're using.
>
> --Nick
>