[j-nsp] RES: Bad TCP Session Performance
Luiz Antonio F. da Silva
luiz at rnp.br
Fri Mar 31 14:13:08 EST 2006
There are two problems, one about TCP flow control mechanism, the sliding
window, and the other with AIMD (Additive Increase Multiplicative Decrease)
the TCP congestion control mechanism.
If you have hosts with fine dimensioned buffers you'll eliminate the first,
but the second is more complex, because of links and routers in the way the
traffic flows. If your traffic experiences a packet drop, because of a
saturated link in the way, the congestion window will be reduced to half.
There is a rule of thumb that says core routers interfaces need to have
sufficient buffer to receive 250ms of full bandwidth, but how to know if all
the routers have well dimensioned buffers?
Tests with iperf or netperf will be very effective to understand if you have
problem in network or application. These tools have options to set buffers
directly without SO tuning.
De: Rafal Szarecki [mailto:rszarecki at gmail.com]
Enviada em: sexta-feira, 31 de março de 2006 15:35
Para: Luiz Antonio F. da Silva
Cc: juniper-nsp at puck.nether.net
Assunto: Re: [j-nsp] RES: Bad TCP Session Performance
I guess Luiz point behavior of TCP sliding window. What is negotiated
max window size?
The largest window should help.
For testing, you can use free tool like iperf.
Rafal Szarecki JNCIE
Luiz Antonio F. da Silva napisal(a):
> Hi Jens,
> There is a known problem in TCP congestion control mechanism, where links
> with high product bandwidth * delay aren't well used by a TCP connection.
> What delay do you observe in your link?
> Luiz Antonio
> juniper-nsp mailing list juniper-nsp at puck.nether.net
More information about the juniper-nsp