[Bloat] [Cerowrt-devel] FQ_Codel lwn draft article review

[Greg White]

Good point.  I wrote that without double checking my code.  What I
should've said was:

Latency is min_latency + 60ms.

This is under the assumption of a fixed 60ms dejitter buffer that is
prescient as to what the min_latency will be.

-Greg



On 11/28/12 12:45 PM, "Simon Barber" <simon at superduper.net> wrote:

>Not sure why you would use mean latency - for VoIP the thing that
>matters is maximum latency - if some packets are late, then you don't
>hear them. More precisely a small packet loss rate is OK, so you should
>use something like the 99th percentile of latency. In my last jitter
>buffer implementation I ran an estimator for 99th percentile latency,
>and added 5ms.
>
>That said, most jitter buffers in use are poor implementations, and
>introduce more latency than necessary.
>
>Simon
>
>
>On 11/28/2012 11:21 AM, Greg White wrote:
>> Jitter on its own is not useful for estimating VoIP quality.
>>
>> For (c), I've used:
>>
>> 	
>> 		
>> 		
>> 	
>> 	
>> 		
>> 			
>> 				
>> 					Cole, Robert G., and Joshua H. Rosenbluth. "Voice over IP
>>performance
>> monitoring." ACM SIGCOMM Computer Communication Review 31, no. 2 (2001):
>> 9-24.
>>
>> Which I generally simplify to:
>> 	
>> 		
>> 		
>> 	
>> 	
>> 		
>> 			
>> 				
>> 					R= 94.2 - 0.024*Latency - 0.11*max(0,Latency-177.3 ms) - 30*log(1 +
>> 15*Loss).
>>
>> 				
>> 			
>> 		
>> 	
>> where Loss is mean packet loss rate (counting packets with latency >
>> (min_latency + 60ms) as lost), Latency is mean latency for packets not
>> counted as lost.
>>
>>
>> 				
>> 			
>> 		
>> 	
>> You can then translate R into an estimate of MOS using Eq.1 in the above
>> reference.
>>
>> -Greg
>>
>>
>> On 11/27/12 5:51 PM, "Toke Høiland-Jørgensen" <toke at toke.dk> wrote:
>>
>>> Oliver Hohlfeld
>>> <oliver at net.t-labs.tu-berlin.de> writes:
>>>
>>>> The jitter measurements you have in mind will give you an idea on the
>>>> jitter specific to the chosen traffic scenario, nothing more --- and
>>>> in particular not the VoIP quality (although low vs. high jitter could
>>>> /indicate/ certain /possible/ quality degradations).
>>>
>>> Well no, in this sense the only "real" test for voip quality is picking
>>> up the (soft)phone and talking to someone. However, since the context
>>> here is automated measuring tools (preferably generating solid
>>> quantitative, comparable data), that is hardly feasible.
>>>
>>> I guess the goal of a comprehensive testing suite is to gather as many
>>> indicators of quality degradations (in the widest possible sense) as
>>> possible and testing for them under a variety of traffic conditions. I
>>> am by no means an expert on VoIP, but someone suggested measuring
>>>jitter
>>> could be useful, and I've proposed a possible way to do that (using
>>> iperf udp flows at a low-ish bandwidth).
>>>
>>> Since for the purpose of this particular discussion I seem to be in the
>>> test tool building business (at least for the time being), what I
>>>really
>>> need before going forward with this is someone to comment on (a) if
>>> using iperf udp flows is a valid way to measure jitter, (b) if
>>>measuring
>>> jitter is actually something someone wants to do and (c) if there are
>>> other tests that would be useful for testing VoIP (or general)
>>> conditions instead of / in addition to the jitter measurements.
>>>
>>> So far I don't have an answer to (a), only negative answers to (b) and
>>> nothing concrete for (c). So for the time being I'm shelving the idea,
>>> and will just note that it seems quite feasible to return to it should
>>> someone change their mind on (b) :)
>>>
>>>
>>> -Toke
>>>
>>> --
>>> Toke Høiland-Jørgensen
>>> toke at toke.dk
>>
>> _______________________________________________
>> Bloat mailing list
>> Bloat at lists.bufferbloat.net
>> https://lists.bufferbloat.net/listinfo/bloat
>>