[e2e] Are we doing sliding window in the Internet?
pingali at ISI.EDU
Tue Jan 2 12:29:55 PST 2007
Detlef Bosau wrote:
> Venkata Pingali wrote:
>> Server end (i.e, end that has large
>> amount of data to transfer):
>> - Most connections are short (90% < 1sec)
> Do you have any knowledge of the number of "rounds" the TCP connection
> has seen? A couple of years ago I saw some similar result (don´t no the
> source at the moment) where 90 % of connections consist of not more than
> 20 packets.
Our sample shows that 94% of connections
have < 20 packets - when observed from the
Number of Packets Percentile of Connections
I have included the new graph and generated pdfs.
> Now, consider the initial slowstart, IIRC we start with 2 MSS (?) then
> we have:
> Round CWND
> 1 2
> 2 4
> 3 8
> total of 14 packets up to now
> 4 16
> total of 24 packets up to now,
> thus many flows will finisch before the end of the fourth round which
> would correspond to a CWND of about 6 kByte, 1500 byte MSS assumed.
> In short words: Quite a few connections are finished before the end of
> the fist slow start period.
> Does this match your observations?
About 90-95% finished before slow start
completed - often within the first two round
trips. About 3-4% of connections lasted for a
long time (several secs - minutes). But there
is an interesting category of connections that
last beyond the slow start but not for very
long. These connections, it turns, carry a large
chunk of the data (40+%) and most of the time in
these connections is spent in slow start.
>> - MaxCwnd is < 5KB in > 80% of cases
>> - MaxRTT is distributed almost uniformly
>> in the 0-400ms range.
>> Client end (i.e., the end receiving data):
>> - ~ 90% of connections see MaxCwnd < 5KB
>> - < 1% connections see MaxCwnd > 10KB
>> - 90% of connections have MaxRTT < 100ms
> Oh, I love it :-)
> Last year I had a long argument with someone who told me about the
> benefits of window scaling :-) He talked about extremely large CWNDs by
> several dozens or hundreds of MByte :-)
Dont know if it is correct to extrapolate from the
same that we have but the MaxCwnd graph seems to
plateau as the connection length increases (bytes
> O.k., that´s a different story because we are talking about greedy
> sources than. However, if that colleague was the only one to activate
> window scaling while surfing from the US and A to good ol´ Europe and
> Cisco et al. had buried hundreds of megabytes of useless queue memory in
> their hardware *blush* this guy perhaps filled the queues the first time
> ever, following the good old paradigm: "Keep the queue full" and that
> way of course outperformed his competitors hopelessly ;-)
>> There are some problems with the data:
>> - limited scenarios (web based)
>> - small sample sizes (21K for server, 150K
>> for client)
>> - the website has non-standard distribution
>> of file types and sizes
> At least it exists. And reality is often more convincing than standards.
> Particularly in cases were both disagree.
>> You can find the various graphs here:
> Just a question: Is it possible to export those slides to a common
> readable format like PDF? I don´t have any M$ products in use here and
> when I opten PowerPoint slides with OpenOffice the results are sometimes
> interesting, sometimes surprising, sometimes hopeless, but nearly always
> quite different from what you wrote :-)
More information about the end2end-interest