Skip to contentSkip to conversation

linux/tc vs packetshaper

LARTC Joakim Recht [email protected] 2 participants 3 messages Feb 1, 2001 – Feb 2, 2001

lore
Share & tools
lore

Joakim Recht [email protected]

author
#0 lore raw

<PRE>Hi

I was wondering how a Linuxbox running iproute2 and tc to do bandwidth control on a 20 mbit internet connection compares? A Packeteer PacketShaper 4500 costs about $20,000, and a Linuxbox can be set up for much less, but can it be used for a network with some 1000 users without too much trouble?
Following that question, how is the latency affected by a Linux router? The PacketShaper should add no more than 50 msec pr packet, how does Linux compare to this?

If someone has experience with a machine connected to an internetconnection with the same bandwidth, I would like to hear about it very much.

Now, I realize that the Packeteer thing is much easier to configure, and also provides some pretty extensive statistical information, but that is not necessary for me.

yours,

Signature · 8 lines
Joakim Recht
Tlf. 20 85 54 77
Email <A HREF="mailto:[email protected]">[email protected]</A>
WWW <A HREF="http://www.braindump.dk">http://www.braindump.dk</A> / <A HREF="http://www.compuclub.dk">http://www.compuclub.dk</A>



</PRE>
#1 lore raw

<PRE>On Thu, Feb 01, 2001 at 07:07:19PM +0100, Joakim Recht wrote:
&gt;<i> Hi
</I>&gt;<i>
</I>&gt;<i> I was wondering how a Linuxbox running iproute2 and tc to do bandwidth </I>&gt;<i> control on a 20 mbit internet connection compares? A Packeteer </I>&gt;<i> PacketShaper 4500 costs about $20,000, and a Linuxbox can be set up for </I>&gt;<i> much less, but can it be used for a network with some 1000 users without </I>&gt;<i> too much trouble?
</I>
Number of users is not that big an issue. We've filtered 90mbit/s of traffic with Linux, but we currently only shape up to 5mbit/s. However, the sites we filter have &gt;100 http connections per second, which each take a few seconds to complete, so I'd guess that at any one time we have ~300 tcp/ip sessions running.

I bet a 1000 users would not come near to that amount of sessions.

&gt;<i> Following that question, how is the latency affected by a Linux router? </I>&gt;<i> The PacketShaper should add no more than 50 msec pr packet, how does </I>&gt;<i> Linux compare to this?
</I>
Well, AFAIK, Packeteers, which used to be quite evil wrt to corrupting IP options by the way!, 'shape' by fooling around with TCP Window options. That does not induce a lot of latency.

Regular shaping might induce more, but it hasn't been a problem for us so far. As long as you are not actively shaping (ie, remain below your bandwidth ceiling), we don't see *any* additional latency.

&gt;<i> Now, I realize that the Packeteer thing is much easier to configure, and </I>&gt;<i> also provides some pretty extensive statistical information, but that is </I>&gt;<i> not necessary for me.
</I>
We've only configured our shaping solution once, using the excellent CBQ.init script. Takes little time, and then runs on without any problems.

Regards,

bert hubert

Signature · 6 lines
PowerDNS                     Versatile DNS Services  
Trilab                       The Technology People   
'SYN! .. SYN|ACK! .. ACK!' - the mating call of the internet


</PRE>
author
#2 lore raw
In reply to Joakim Recht [email protected]

<PRE>
bert hubert wrote:

&gt;<i>
</I>&gt;<i> Number of users is not that big an issue. We've filtered 90mbit/s of traffic </I>&gt;<i> with Linux, but we currently only shape up to 5mbit/s. However, the sites we </I>&gt;<i> filter have &gt;100 http connections per second, which each take a few seconds </I>&gt;<i> to complete, so I'd guess that at any one time we have ~300 tcp/ip sessions </I>&gt;<i> running.
</I>&gt;<i>
</I>&gt;<i> I bet a 1000 users would not come near to that amount of sessions. </I>
Well, I've had to increase /proc/sys/net/ipv4/ip_conntrack_max from 16k
to a larger value on a network with 300 users, so something tells me
that quite a lot of connections are open.

&gt;<i>
</I>&gt;<i> Regular shaping might induce more, but it hasn't been a problem for us so </I>&gt;<i> far. As long as you are not actively shaping (ie, remain below your </I>&gt;<i> bandwidth ceiling), we don't see *any* additional latency. </I>
This is the problem for me... The connection is to be distributed between some 8 college dorms, and the users here use quite a lot of bandwidth right now, as we are connected to a university connection. However, we are about to buy our own, which means a pretty crowded 20 mbit connection, and as each dorm should be guaranteed an amount of bandwidth, shaping could occur very often, at least at peak hours in the afternoon/evening.

&gt;<i> We've only configured our shaping solution once, using the excellent </I>&gt;<i> CBQ.init script. Takes little time, and then runs on without any problems. </I>&gt;<i>
</I>&gt;<i> Regards,
</I>&gt;<i>
</I>&gt;<i> bert hubert
</I>&gt;<i>
</I>Thanks for your reply,

Signature · 8 lines
Joakim Recht
Tlf. 20 85 54 77
Email <A HREF="mailto:[email protected]">[email protected]</A>
WWW <A HREF="http://www.braindump.dk">http://www.braindump.dk</A> / <A HREF="http://www.compuclub.dk">http://www.compuclub.dk</A>



</PRE>