[comm] Throughput on a lossy link
David Halstead
dhalstea at nrao.edu
Thu May 27 19:03:27 EDT 2010
James,
Gene and I were mulling this over and remembered reading about the
bandwidth wall you hit with higher latency links with even minor packet
loss.
Baseline: To achieve a gigabit/sec throughput on a 1500MTU link with
40ms RTT you need a packet loss of less than 0.00001%!!
A loss of 0.1% gives a max throughput of ~9Mbps for a single TCP stream,
even for a 10Gig link of the same latency.
For more info see: http://sd.wareonearth.com/~phil/issues.html and
Not that this changes the plan to fix the local link to NMT, but that
problem MAY also explain the 99.9% bandwidth loss to CV due to the
1/(rtt*sqrt(loss)) term.
David
James Robnett wrote:
> Just did the network tests with NMT. Off hand it looks like
> there's a bad fiber somewhere between us since the tests ruled out
> the transceiver.
>
> We're getting no where near the throughput we should be getting over
> that 1gbit link to NMT and UNM.
>
> This *doesn't* explain the poor performance to CV, in the AOC to
> NMT/UNM case we're missing many tens of percent bandwidth. In the
> AOC to CV case we're missing 99.9% of the bandwidth.
>
> James
>
> Gene Runion wrote:
>
>> https://staff.nrao.edu/wiki/bin/view/Ccs/MonthlyCommunicationMeetings
>>
>> There will be a communication meeting this Wednesday at 17:30UT/13:30ET/11:30MT.
>>
>> Locations: CV-conf (CV-209), GB-conf2 (GB-241), SO-conf (AOC-317)
>>
>> Gene Runion
>> 434-296-0327
>>
>>
>>
>> _______________________________________________
>> Comm mailing list
>> Comm at listmgr.cv.nrao.edu
>> http://listmgr.cv.nrao.edu/mailman/listinfo/comm
>>
> _______________________________________________
> Comm mailing list
> Comm at listmgr.cv.nrao.edu
> http://listmgr.cv.nrao.edu/mailman/listinfo/comm
>
More information about the Comm
mailing list