[LARTC] High Latency With Tiered Queues

[LARTC] High Latency With Tiered Queues

[ISN Support Staff]

Hello,

  I'm trying to setup what I thought would be a fairly basic tiered
shaping system.  I have a 6mbit (768kbps) link coming into my eth1
device, with my LAN IPs on the eth0 device.  I want to limit outgoing
traffic so that certain IPs are limited to 400kbps, with 3 classes under
that 400k so certain machines get prioritized (main servers in 1:21,
other servers in 1:22, workstations in 1:23)

  The problem is that when I turn this on, my packet latency jumps up by
50 to 100 times the normal rate.  I go from 10-20 ms ping times to
500-1600ms!  I've tried putting SFQ qdiscs under the classes, but that
makes no difference.

  I'm sure there is just some tuning parameter I'm not setting
correctly, but can somebody clue me in to what I'm doing wrong?  Or is
HTB just the wrong scheduler to be using here?  I tried CBQ, but I can't
get the tiers to work ( I keep getting RTNETLINK answers: Invalid
argument)  I'm currently using a single tiered CBQ solution, but it
really doesn't fit my needs.

Here's the full script:
-----------------------
qdisc add dev eth1 root handle 1: htb default 10
class add dev eth1 parent 1: classid 1:1 htb rate 768kbps
class add dev eth1 parent 1: classid 1:10 htb rate 250kbps ceil 768kbps
burst 6k
class add dev eth1 parent 1: classid 1:20 htb rate 100kbps ceil 400kbps
burst 6k
class add dev eth1 parent 1:20 classid 1:21 htb rate 75kbps ceil 375kbps
burst 6k
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.35 flowid 1:21
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.29 flowid 1:21
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.28 flowid 1:21
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.16 flowid 1:21
class add dev eth1 parent 1:20 classid 1:22 htb rate 15kbps ceil 350kbps
burst 6k
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.30 flowid 1:22
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.31 flowid 1:22
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.32 flowid 1:22
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.36 flowid 1:22
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.39 flowid 1:22
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.40 flowid 1:22
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.47 flowid 1:22
class add dev eth1 parent 1:20 classid 1:23 htb rate 10kbps ceil 200kbps
burst 6k
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.33 flowid 1:23
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.34 flowid 1:23
filter add dev eth1 parent 1: protocol ip prio 16 u32 match ip src
192.168.1.37 flowid 1:23


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

Re: [LARTC] High Latency With Tiered Queues

[Andy Furniss]

ISN Support Staff wrote:
> Hello,
> 
>   I'm trying to setup what I thought would be a fairly basic tiered
> shaping system.  I have a 6mbit (768kbps) link coming into my eth1
> device, with my LAN IPs on the eth0 device.  I want to limit outgoing
> traffic so that certain IPs are limited to 400kbps, with 3 classes under
> that 400k so certain machines get prioritized (main servers in 1:21,
> other servers in 1:22, workstations in 1:23)
> 
>   The problem is that when I turn this on, my packet latency jumps up by
> 50 to 100 times the normal rate.  I go from 10-20 ms ping times to
> 500-1600ms!  I've tried putting SFQ qdiscs under the classes, but that
> makes no difference.
> 
>   I'm sure there is just some tuning parameter I'm not setting
> correctly, but can somebody clue me in to what I'm doing wrong?  Or is
> HTB just the wrong scheduler to be using here?  I tried CBQ, but I can't
> get the tiers to work ( I keep getting RTNETLINK answers: Invalid
> argument)  I'm currently using a single tiered CBQ solution, but it
> really doesn't fit my needs.
> 
> Here's the full script:
> -----------------------
> qdisc add dev eth1 root handle 1: htb default 10

Using htb default will send arp to a really long (in this case) 
backlogged queue, which could cause problems.

Generally you need to classify low latency traffic to different class 
than bulk if you want it to stay low latency.

I am suprised sfq didn't help - maybe you also need to back off the 
ceils to allow for overheads on the quoted link speed.

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