|
|
comp.protocols.tcp-ip
Discussion complète de l'article :
WAN packet loss
|
Article :
21813
Date :
08-10-2005
From :
Alan Strassberg
Sujet :
WAN packet loss
What's 'reasonable' packet loss on a WAN for a global enterprise?
By reasonable I mean normal loss that doesn't impact apps (I
realize VoIP has it's own set). Is 1% typical? Less than 2.5%
acceptable? Where does tcp start losing it with retransmissions?
alan
Posez vos questions, réponses et remarques sur
les forums de FrameIP
|
|
Article :
21817
Date :
09-10-2005
From :
stanislav shalunov
Sujet :
Re: WAN packet loss
paleale@bolt.sonic.net (Alan Strassberg) writes:
> What's 'reasonable' packet loss on a WAN for a global enterprise?
Every sane fiber backbone currently in operation has no congestive
packet loss. Non-congestive packet loss can be on the order of 1e-8
before one starts to consider it a problem; it very rarely gets better
than 1e-10 (BER in fiber is bounded from below by a function of
physical light loss).
> By reasonable I mean normal loss that doesn't impact apps
Standard TCP has throughput that's limited by approximately
MSS/(RTT*sqrt(loss)). Therefore, if your MSS is 1460B and RTT 70ms
(typical for a cross-continent path), you need no more no more than
0.0003% packet loss to get 100Mb/s or 0.03% to get 10Mb/s. Note that,
because of that square root, every doubling of required throughput
means dividing permissible packet loss by four.
> Is 1% typical? Less than 2.5% acceptable?
No and no. Not by a long shot. This is *way* above normal levels.
> Where does tcp start losing it with retransmissions?
At about 5% you start getting broken connections. At 10%, you have an
essentially unusable network as far as any interactive performance
with TCP is concerned.
[rearranged]
> (I realize VoIP has it's own set).
VoIP with decent codecs is actually quite resilient, easily tolerating
packet loss. Delay, however, kills the conversation quality
regardless of codec.
--
Stanislav Shalunov http://www.internet2.edu/~shalunov/
Al your Qaeda are belong to US.
Posez vos questions, réponses et remarques sur
les forums de FrameIP
|
|
Article :
21818
Date :
09-10-2005
From :
Walter Roberson
Sujet :
Re: WAN packet loss
In article <863bnbxw37.fsf@abel.internet2.edu>,
stanislav shalunov wrote:
:Standard TCP has throughput that's limited by approximately
:MSS/(RTT*sqrt(loss)). Therefore, if your MSS is 1460B and RTT 70ms
:(typical for a cross-continent path), you need no more no more than
:0.0003% packet loss to get 100Mb/s or 0.03% to get 10Mb/s. Note that,
:because of that square root, every doubling of required throughput
:means dividing permissible packet loss by four.
Your formula implies that TCP throughput approaches infinity
as the loss approaches 0, and postulates a limit of just over
1 terabit per second for a 1 ms RTT with the lower limit BER
you state for fibre (1E-10).
The implication is also that for 499 ms RTT, and 7.21 Gbps,
the loss could be no more than 1E-11, 10 times lower than the
BER for fibre that you cite. It would appear that either your
formula or your cited BER are inconsistant with the claims of
the University of Tokyo....
http://data-reservoir.adm.s.u-tokyo.ac.jp/press/press.html
--
These .signatures are sold by volume, and not by weight.
Posez vos questions, réponses et remarques sur
les forums de FrameIP
|
|
Article :
21830
Date :
10-10-2005
From :
stanislav shalunov
Sujet :
Re: WAN packet loss
roberson@ibd.nrc-cnrc.gc.ca (Walter Roberson) writes:
> Your formula
I might wish, but it isn't mine; it's Matt Mathis's.
> implies that TCP throughput approaches infinity as the loss
> approaches 0,
Yes.
> and postulates a limit of just over 1 terabit per second for a 1 ms
> RTT with the lower limit BER you state for fibre (1E-10).
Not sure why you mention this, but perhaps your point is that terabit
performance does not happen. If that is, indeed, what you're saying,
then this is simply a demonstration of the fact that, at 1ms RTT,
conventional TCP will experience losses not related to well-maintained
fiber errors. If no other brokenness comes up, then it'll be
congestive loss (whether induced by the flow in question or by other
flows).
> The implication is also that for 499 ms RTT, and 7.21 Gbps,
> the loss could be no more than 1E-11, 10 times lower than the
> BER for fibre that you cite.
If you make the fiber squeaky clean and spit on the connectors
lovingly, you might even observe another order of magnitude
improvements. Measurement beyond that point gets uncertain enough
that it's difficult to say where the limit is. The OP was asking
about normal WAN packet loss rates, though, not about what can be done
by good network engineer for a record.
> It would appear that either your formula or your cited BER are
> inconsistant with the claims of the University of Tokyo....
>
> http://data-reservoir.adm.s.u-tokyo.ac.jp/press/press.html
The BER rate for R&E networks is quite good in general, yet there's
always extra work involved with getting the network extra-clean for
this kind of records.
For the entire history of the Land Speed Record competition, you might
take a look at http://lsr.internet2.edu/history.html
The current bottleneck isn't packet loss rate, by the way, but simply
the network card capacity (which has the tendency to drive the
competition towards longer and longer paths).
--
Stanislav Shalunov http://www.internet2.edu/~shalunov/
This message is designed to be viewed with 0.06479891g of NaCl.
Posez vos questions, réponses et remarques sur
les forums de FrameIP
|
|
|
