[quagga-users 12558] Re: bgp - quagga-0.99.18

Ingo Flaschberger if at xip.at
Tue Nov 8 17:57:43 GMT 2011


>> as Nick Hilliard outlined the size and speed of a buffer makes the
>> difference.
>> bufferbloat does not tell which size / speed is bad -
>
> Jim Gettys gives a back of the envelope formula for the optimal buffer
> size in his "bufferbloat" talk[1]:
>
>    delay x bandwidth x ?#flows
>
> The optimal buffer size for a single flow is delay x bandwidth (this is
> well known). I'm not sure where he gets the sqrt from (intuitively it is
> clear that the buffer size must not increase linearily with the number
> of flows, but how much less it must be I don't know).
>
> Less is bad because you waste time waiting for ACKs.
> More is bad because you get excessive lag on congested links and TCP
> won't throttle back.

         +--------------------+
input ->|    1 mbit queue    |-> output
         +--------------------+
                             ^
                             |
                          small packet input

tcp needs some backpresure to scale down, better than dropping packets 
with a too short queue. delay ~40-100ms.
the key is the "fast path" for small packets < 128byte, mainly tcp acks.
a link can now become congested in both directions - but without slowing 
acks.

but nearly all qos / rate-limiting implementations use this.

and yes - offtopic.

Kind regards,
 	Ingo Flaschberger


More information about the Quagga-users mailing list