Brian J. Murrell wrote: > > decrease throughput, cause the connections to > > stall > > Well they only stall as long as it takes for TCP to resend the packet, > which should not be a huge amount of time.
Unless there is a artificial high packet delay such as a badly tuned queue, in which case the TCP retransmission algorithm gets a bit screwed up.. To recap: * To shape bandwidth, drop packets. Applications will recover. * To ensure a link is best utilized, queue packets when reaching the limit. But make sure this queue isn't too large counted in link time (generally not more than a few ms can be recommended) or else the queue will screw up RTT timing of the link, and unless a really good queue policy is used a large queue easily allows for a single session to monopolize the available bandwidth. * To ensure the traffic gets a chance to nicely deal with the bandwidth limitation, make the droppoint smooth by using algorithms such as RED.