LARTC - Sep 2006 - HTB root rate allowing to much of a burst

Please see below

Jon Flechsenhaar
Boeing WNW Team
Network Services
(714)-762-1231
202-E7

-----Original Message-----
From: Flechsenhaar, Jon J 
Sent: Wednesday, September 27, 2006 10:30 AM
To: lartc@mailman.ds9a.nl
Subject: FW: [LARTC] 2.6.14 - HTB/SFQ QoS broken?

 Please see below


Jon Flechsenhaar
Boeing WNW Team
Network Services
(714)-762-1231
202-E7

-----Original Message-----
From: Flechsenhaar, Jon J
Sent: Wednesday, September 27, 2006 9:35 AM
To: lartc@mailman.ds9a.nl
Subject: [LARTC] 2.6.14 - HTB/SFQ QoS broken?

Jody:

I don''t know if you answer basic HTB questions but I''l try
anyway.  I''m
implementing AF and EF with a root rate of 500 kbps (kilo bits per
second).  I am using DSMARK to classify packets.  I''m implementing this
on two different hardware sets PPC and x86 with the same kernel options
configured.  

Problem:

My root rate is going about 75 kbps over its 500kbps limit (ceil 500).
I believe this is because of burst and cburst settings.  I am using
default calculations now, on all classes.   Is there a formula to use in
tweaking these
values?  I would like my rates to follow the assigned rate more closely.


I am generating traffic with a hardware traffic generator.  Starting at
300 kbps stepping by 100 kbps up to 1mbps.  Each step lasts for 60 sec.


My config script is attached.  Thanks.  

Jon

 <<diffserv9-26-gred>> 

Jon Flechsenhaar
Boeing WNW Team
Network Services
(714)-762-1231
202-E7



_______________________________________________
LARTC mailing list
LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Flechsenhaar, Jon J wrote:
> 
> I don''t know if you answer basic HTB questions but I''l
try anyway.  I''m
> implementing AF and EF with a root rate of 500 kbps (kilo bits per
> second).  I am using DSMARK to classify packets.  I''m implementing
this
> on two different hardware sets PPC and x86 with the same kernel options
> configured.  
> 
> Problem:
> 
> My root rate is going about 75 kbps over its 500kbps limit (ceil 500).
How are you measuring - htb and rate estimators see packets on eth an 
iplen + 14.
> I believe this is because of burst and cburst settings.  I am using
> default calculations now, on all classes.   Is there a formula to use in
> tweaking these
> values?  I would like my rates to follow the assigned rate more closely.
Hz and a compile time header change - HYSTERESIS 0 in 
net/sched/sched_htb.c also affect accuracy.
> 
> I am generating traffic with a hardware traffic generator.  Starting at
> 300 kbps stepping by 100 kbps up to 1mbps.  Each step lasts for 60 sec.
Could give different results to normal tcp I suppose.
> 
> 
> My config script is attached.  Thanks.  
tbf latency 0.01 on 256kbit is too short for a buffer IIRC it gets 
rounded up so at least mtu packets get through.

I get an error on your script with iproute2-ss060323 and 2.6.17.11/7

++ /usr/sbin/tc qdisc change dev eth0 parent 2:20 gred DP 3 limit 60kb 
min 10kb max 20kb avpkt 1500 burst 9 bandwidth 500kbit probability 0.06 
prio 4
RTNETLINK answers: Invalid argument

after each DP 3 on each gred.

Andy.