[Tmrg] Queue size - Towards a Common TCP Evaluation Suite

[quetchen at caltech.edu (Tom Quetchenbach)]

I tried some experiments with my dial-up connection yesterday. I had to
download several files at once to reach what seemed to be close to a
maximum delay.

Here is a summary of my results:

Uncongested:
min: 140ms, max: 171ms, mean: 154ms

Congested (six large background flows from kernel.org servers):
min: 3936ms, max: 6780ms, mean: 5407ms

So, max(congested) - min(uncongested) = 6640 ms

My modem reported a connection speed of 54.6 Kbit/s, so

6.650 s * 54.6 Kbit/s / 8 = 45 Kbyte/s
45 Kbyte/s / 1.5 Kbyte/packet = 30 packets

This was around 10-11 AM PST on 2008/10/03, using Windows XP
Professional (service pack 3). The background traffic was between two
and six large files from ftp://ftp2.kernel.org, http://kernel.org, and
http://mirrors.kernel.org. The ISP was AT&T in Pasadena, CA.

Here are my raw data:

http://wil-ns.cs.caltech.edu/~quetchen/dialup-tests/ping-output/

And, in the interest of putting off real work, here is a rough plot of
ping RTT vs. time:

http://wil-ns.cs.caltech.edu/~quetchen/dialup-tests/ping_rtt.png

Would it be worth re-running the test with smaller packets, to see if
the queue size in this case is specified in bytes or packets? I think I
can convince Windows to change its MTU. I was also planning on testing
with hping2 (which uses TCP SYN packets instead of ICMP echos) and
comparing the results. I'll do this sometime this weekend or Monday.

-Tom

Tom Quetchenbach wrote:
> My ISP gives me dial-up access as a backup to my DSL, so I'll try to
> play around with it at some point.
> 
> -Tom
> 
> Lachlan Andrew wrote:
>> Thanks Stefan.
>>
>> Those numbers are interesting.  I'm surprised that there was 8s delay
>> when congested.  I'm wondering if  ping  packets are treated
>> differently.  (Many systems give ICMP packets lower priority.) Still,
>> 35 packets sounds a reasonable buffer size.
>>
>> Does anyone else on the list have any data to support or contradict
>> this?  My parents-in-law use dial-up, so I'll try to check their
>> connection soon.
>>
>> Cheers,
>> Lachlan
>>
>> 2008/10/2 Stefan Hirschmann <krasnoj at gmx.at>:
>>> Greeting Andrew and all other readers,
>>>
>>>> Lachlan Andrew wrote:
>>>>> 2008/7/30 Stefan Hirschmann <krasnoj at gmx.at>:
>>>> Greetings Stefan,
>>>>
>>>> Thanks for your interest in the test suite.  I apologise for the long
>>>> delay in getting back to you.
>>> I apologize for this long delay too. But it was not easy to find anyone
>>> with a 56K POTS modem still in use.
>>>
>>>
>>>>> In the "Common TCP Evaluation Suite draft-irtf-tmrg-tests-00" there is the section:
>>>>> "3.2. Delay/throughput tradeoff as function of queue size"
>>>>> describing the buffer sizes of the routers, but only for the access link scenario.
>>>>>
>>>>> I wanted to extend the values to the other scenarios and noticed a problem with it.
>>>>> The BDP of the Dial-Up Link scenario is 64Kbps * 0.1 s / 8 = 0.8 KByte -> 0.8 / 1.5 = 0,53 packets.
>>>>>
>>>>> So even if I use the BDP the value is much too small. A rounding to one is IMHO also not realistic. What value should be used as a minimum buffer size and why?
>>>> The Dial-Up scenario is there partly for POTS modems, and partly for
>>>> GPRS.  You should find out the buffer size used by either one of those
>>>> (and then it would be great to post it to the list!).
>>>>
>>>> If you have access to a dial-up connection, you could try to measure
>>>> the buffer size:  Ping the next-hop node with an idle link, and then
>>>> while downloading something large.  The difference in RTTs will give a
>>>> good estimate of the buffer size.
>>> OK I have done it. The test were made:
>>> DATE: 2008/10/01 around 19:30
>>> Used 56K POTS Provider: Tele2 Austria
>>> Operating System: Windows XP Media Centre Edition
>>> Large background traffic: A linux kernel image from ftp://ftp2.kernel.org
>>>
>>> The exact testprotocol is at the end of the email.
>>> The most important datas are:
>>> uncongested:
>>> Minimum = 134ms, Maximum = 148ms, Mean = 141ms
>>>
>>> congested:
>>>  Minimum = 5963ms, Maximum = 8541ms, Mittelwert = 7177ms
>>>
>>> The correct formula should be:
>>> max(queuing time) = max(congested) - min(uncongested)
>>> 8407 ms = 8541 ms - 134 ms
>>>
>>> 56 KBit/s is 7 KByte/s. 6 KByte/s is a realistic value for the real
>>> usable value. In this case:
>>> time * bandwidth = amount of data
>>> 8,541 s * 6 KByte/s = 51,246 KByte
>>>
>>> If you say, that the packetsize is 1,5 KByte than:
>>> 51,246 KByte / 1.5 KByte =  34,164
>>>
>>> So 35 is the Queuesize in packets.
>>>
>>>
>>>
>>> Cheers Stefan
>>>
>>>
>>>
>>> Now the complete console log (was a German Windows vesion):
>>> ===============================================================================================
>>> Microsoft Windows XP [Version 5.1.2600]
>>> (C) Copyright 1985-2001 Microsoft Corp.
>>>
>>> C:\Dokumente und Einstellungen\Leo>ping www.google.at
>>>
>>> Ping www.l.google.com [209.85.129.147] mit 32 Bytes Daten:
>>>
>>> Antwort von 209.85.129.147: Bytes=32 Zeit=148ms TTL=244
>>> Antwort von 209.85.129.147: Bytes=32 Zeit=146ms TTL=244
>>> Antwort von 209.85.129.147: Bytes=32 Zeit=136ms TTL=244
>>> Antwort von 209.85.129.147: Bytes=32 Zeit=134ms TTL=244
>>>
>>> Ping-Statistik f?r 209.85.129.147:
>>>    Pakete: Gesendet = 4, Empfangen = 4, Verloren = 0 (0% Verlust),
>>> Ca. Zeitangaben in Millisek.:
>>>    Minimum = 134ms, Maximum = 148ms, Mittelwert = 141ms
>>>
>>>
>>> C:\Dokumente und Einstellungen\Leo>ping -w 9999 www.google.at
>>>
>>> Ping www.l.google.com [209.85.129.104] mit 32 Bytes Daten:
>>>
>>> Antwort von 209.85.129.104: Bytes=32 Zeit=7027ms TTL=244
>>> Zeit?berschreitung der Anforderung.
>>> Antwort von 209.85.129.104: Bytes=32 Zeit=8541ms TTL=244
>>> Antwort von 209.85.129.104: Bytes=32 Zeit=5963ms TTL=244
>>>
>>> Ping-Statistik f?r 209.85.129.104:
>>>    Pakete: Gesendet = 4, Empfangen = 3, Verloren = 1 (25% Verlust),
>>> Ca. Zeitangaben in Millisek.:
>>>    Minimum = 5963ms, Maximum = 8541ms, Mittelwert = 7177ms
>>>
>>> C:\Dokumente und Einstellungen\Leo>
>>>
>>
>>
> 

-- 
/* Tom Quetchenbach
 * WAN-in-Lab / Netlab, Dept of Computer Science, Caltech
 * 1200 E California Blvd, MC 256-80, Pasadena CA 91125
 * Lab: [phone]  ||  Cell: [phone]
 */