Re: [ippm] Comments on draft-ietf-ippm-framework-compagg-00

[Anura Jayasumana <[e-mail]>]

Hi Reza,
You hit the key point I  tried to avoid as the answer is not simple.
The value is relative,  and depends on intermediate link speeds, current 
congestion etc. 
20 ms would be small in a slow bit-rate network, or if going through
a narrow bottleneck, while it may be high in a fast network - say a 
100Mbps LAN.   
One of the main causes of correlation is due to later packets of a 
stream having to wait (e.g. in a queue) while prior packets in the same 
stream are being processed.  

Although we were not directly addressing the issue being disussed,  a 
paper at ICON 2004 -  
http://www.engr.colostate.edu/ece/faculty/jayasumana/pdf/conferences/icon_04_np.pdf
contains measurements and a discussion illustrating  this problem.

Cheers!
Anura


Fardid, Reza wrote:

>Anura,
>
>Can you please quantify small and large inter-packet gaps, in the
>context of delay?
>Do you consider 20 msec, the frame length of typical VoIP packets,
>small?
>
>  
>
>Regards,
>Reza Fardid
>Covad Communications
>
>-----Original Message-----
>From: Anura Jayasumana [mailto:[e-mail]] 
>Sent: Tuesday, March 21, 2006 6:45 PM
>To: Philip F. Chimento Jr.
>Cc: ippm
>Subject: RE: [ippm] Comments on draft-ietf-ippm-framework-compagg-00
>
>  
>
>>===== Original Message From "Philip F. Chimento Jr."
>>    
>>
><[e-mail]> 
>=====
>  
>
>>Hi Anura:
>>You can add means, whether the relevant RVs are dependent or not, so
>>average delays add.
>>Higher order statistics, as you point out, do not add simply.
>>Convolution ASSUMES that the RVs are independent.
>>
>>
>>Regards,
>>Phil Chimento
>>    
>>
>
>Phil,
>Yes, you are right.  Based on very limited results we have for delay, it
>looks 
>like the delays are independent if the inter-packet gaps of the stream
>is 
>large; when ipg is small, such as with a burst, the delays appear to be
>highly 
>correlated. This is when convolution breaks down. 
>Cheers!
>AJ
>
>  
>
>>On Mar 21, 2006, at 12:31, Anura Jayasumana wrote:
>>
>>    
>>
>>>Henk Uijterwaal wrote:
>>>      
>>>
>>>>Phil,
>>>>
>>>>The problem that I have with this section is that it doesn't give me
>>>>very strong or concrete reasons for why I want to combine metrics.
>>>>
>>>>Suppose I have a path A-B-C.  I'm interested in the delays between
>>>>them.
>>>>Now, I can measure all 3 (A-B, B-C and A-B-C) separately.
>>>>However, if
>>>>there is a way to add up the delays between A-B and B-C, and get the
>>>>overall delay, then i can see myself some work and diskspace.  In
>>>>general,
>>>>I'd be interested in the minimum set of measurements that still
>>>>        
>>>>
>gives
>  
>
>>>>me information about the entire network.
>>>>
>>>> Section 5.1: The problem here is that piece-wise measurements do
>>>>not necessarily combine. IPDV and re-ordering are examples where one
>>>>subsection of a network can "undo" what a previous one has done.
>>>>Putting these together in a straightforward way doesn't give you the
>>>>right answer.
>>>>        
>>>>
>>>With delays,  if the packet delays in the two networks are not
>>>correlated, it is still possible to measure statistics of  delays
>>>of segments and combine them to provide end-to-end delay
>>>statistics.  When delays are correlated, it becomes necessary to
>>>measure correlation as well.... which would be more complex.
>>>Reordering is somewhat of a harder problem. But, when  captured
>>>using RD (reorder density), you can combine the reordering in
>>>different segments and still come up with end-to-end reordering.
>>>Combination is done using
>>>convolution, which would be the case for delay also.  One thing
>>>going for RD is the fact that reordering in two network segments
>>>are less likely to be correlated compared to delay.  We have
>>>carried out many mesurements on Internet, planet-lab, simulations,
>>>etc., and are yet to see a case where it did not work - for
>>>reordering.
>>>(See http://www.cnrl.colostate.edu/packet_reorder.html )
>>>Cheers!
>>>AJ
>>>
>>>      
>>>
>>>>I agree but can't we just say so?
>>>>
>>>>Henk
>>>>
>>>>
>>>>
>>>>        
>>>>
>-
>

Anura P. Jayasumana <http://www.engr.colostate.edu/%7Eanura>
Professor, Electrical & Computer Engineering
                           and Computer Science
Department of Electrical and Computer Engineering
Colorado State University
Fort Collins, CO 80523
Phone: [phone]
Fax: [phone]
Email: <[e-mail]>

_______________________________________________
ippm mailing list
[e-mail] 
https://www1.ietf.org/mailman/listinfo/ippm