[grobinson]

Can someone how traffic shaping reduces latency rather than simply lowering the amount of load required to cause significant latency? I understand that the aim here is to prevent the router from having to queue packets, but by doing this haven't we just shifted the queue and source at which packets are dropped from the router to the Linux box?

[ploxiln]

The problem is that the router does not drop packets (as soon as a proper traffic shaper does). It queues them up, megabytes at a time. Packets that do make it to their destination take longer and longer to get there (due to waiting in the queue).

Eventually most of the buffered packets are so late that they're considered missing/dropped, though they took bandwidth to send and caused other packets to wait behind them. Then TCP adjusts to a long series of lost packets by dropping the rate dramatically.

TCP isn't supposed to work that way :)

[agshekeloh]

(Note: This link points to my web site, it's my story.)

You can't erase the capacity problems that cause issues like latency. Depending on your gear, however, you can shift them.

In this case, I used traffic shaping to transfer the latency to the customer that generated the traffic.

It let other customers get the normal usage that they'd paid for.

==ml